diff --git a/README.md b/README.md index b229fe3..c315ece 100644 --- a/README.md +++ b/README.md @@ -2,7 +2,7 @@ This repository contains the sources for [de3ug.github.io](https://de3ug.github.io/). All public files live in the `site/` directory and are deployed automatically to GitHub Pages. -The homepage includes sections for publications, projects, work experience and a Personal/Hobbies area with links to travel, astrophotography and racing content. +The homepage includes publications, service and patents in one section, projects, work experience, and a Personal/Hobbies area that also lists writing. ## How the site is structured @@ -10,9 +10,13 @@ The deployed site is intentionally simple: `site/` is the public web root. GitHu The main page is split into a static shell and data-driven content: -- `site/index.html` defines the layout, navigation, theme styling and empty section containers. +- `site/index.html` defines the homepage shell and empty section containers. +- `site/css/site.css` and `site/js/navigation.js` provide the shared sidebar, layout, and theme styling for the homepage, writing pages, and dataset page. - `site/js/render.js` fetches `site/resume.json` in the browser and fills the homepage sections. - `site/js/theme-toggle.js` handles the light/dark theme toggle. +- `site/writing/` contains the writing archive and standalone long-form articles. +- `site/css/writing.css` adds reading and dataset layouts to the common site styles. +- `site/datasets/viper/` contains the dataset landing page, original ZIP, and citation file. - `site/resume.pdf` is the printable one-page resume artifact. - `site/llms.txt` is a short LLM-readable summary artifact that points agents to `resume.json`. - `site/gcmap/` is a standalone static demo app for mapping great-circle flight routes. @@ -24,7 +28,7 @@ The main page is split into a static shell and data-driven content: The JSON has a hierarchy of information: -- It contains the full content inventory, including public profile text, publications, patents, media mentions, experience, hobbies, contact links and background notes. +- It contains the full content inventory, including public profile text, writing metadata, publications, patents, media mentions, experience, hobbies, contact links and background notes. - Most current, public-facing items are rendered into the webpage. - Some older school, internship and early-career details are kept in the JSON for context but hidden from the webpage because they are no longer central to the public portfolio. - Only the most important highlights are included in the one-page resume PDF. @@ -33,6 +37,14 @@ Visibility is controlled with fields such as `hide_from` and `for`. These fields `build.py` reads the same JSON and generates `site/resume.pdf`, keeping the resume concise while preserving a richer content hierarchy for the website and future edits. +Long-form prose does not live in `resume.json`. The `writing` collection stores discoverability metadata such as title, summary, publication date, canonical URL and site-relative path. Each article owns its full content and assets under `site/writing//`; this keeps the résumé data useful as a directory without turning it into a document format. + +Writing appears under Personal/Hobbies on the homepage. The separate `/writing/` archive and article pages use the same sidebar and light/dark controls as the portfolio. The VIPeR article lives at `/writing/twenty-years-of-person-re-identification/`; the previous preview slug redirects there. + +The VIPeR download page is `/datasets/viper/`. Its ZIP is the exact historical v1.0 archive, with checksums and the original citation. The homepage dataset entry is web-only. The essay's citation chart uses manually observed Google Scholar article histograms, stored with source URLs and observation date in `assets/scholar-citations.json`. Run `python3 build_citation_chart.py` with Matplotlib installed to regenerate it. + +The essay's performance scatter plot uses `assets/rank-one-scatter.json`, which records reported results, evaluation settings, dates, and paper sources. Run `python3 build_rank_one_chart.py` with Matplotlib installed to regenerate `assets/rank-one-scatter.png`. The plot distinguishes training settings and is not a controlled leaderboard. + The `last_updated` field in `site/resume.json` is also the source of truth for the "Last updated" text shown on the website, the PDF resume and the generated `llms.txt` summary. When changing profile content or rebuilding artifacts, update this date in the JSON. ## LLM-readable summary diff --git a/build.py b/build.py index 45985d4..faece4d 100644 --- a/build.py +++ b/build.py @@ -107,6 +107,12 @@ def render_llms(r): for s in r['social'] if visible(s, 'web') ) + writing = '\n'.join( + f'- [{entry["title"]}]({entry["url"]}): {entry["summary"]}' + for entry in r.get('writing', []) + if visible(entry, 'web') + ) + writing_section = f'\n## Writing\n\n{writing}' if writing else '' return f"""# {r['name']} @@ -115,6 +121,7 @@ def render_llms(r): {r['about']} Last updated: {r['last_updated']} +{writing_section} Contact: - Email: {r['email']} @@ -214,6 +221,7 @@ def render_html(r, target='resume'):

Selected Publications & Service

{parts_to_html(r['publications_intro']['parts'], target)}

{ul(r['publications'], target)} +

{r['reviewer_service']}

Selected Patents

{r['patents_intro']}

diff --git a/build_citation_chart.py b/build_citation_chart.py new file mode 100644 index 0000000..4265cf0 --- /dev/null +++ b/build_citation_chart.py @@ -0,0 +1,56 @@ +#!/usr/bin/env python3 +"""Render the observed Google Scholar histories used in the VIPeR essay.""" + +import json +from pathlib import Path + +import matplotlib +matplotlib.use('Agg') +import matplotlib.pyplot as plt + +ASSETS = Path(__file__).resolve().parent / 'site/writing/twenty-years-of-person-re-identification/assets' + + +def main(): + data = json.loads((ASSETS / 'scholar-citations.json').read_text()) + rows = data['annual'] + years = [row['year'] for row in rows] + plt.rcParams.update({'font.family': 'sans-serif', 'font.size': 12}) + fig, ax = plt.subplots(figsize=(10, 5.6), dpi=180) + fig.patch.set_facecolor('white') + ax.set_facecolor('white') + for key, label, color, style in [ + ('eccv2008', 'ECCV 2008', '#11637b', '-'), + ('pets2007', 'PETS 2007', '#b64b24', '--'), + ]: + counts = [row[key] for row in rows] + ax.plot(years[:-1], counts[:-1], color=color, linewidth=2.5, + linestyle=style, label=label) + ax.plot(years[-2:], counts[-2:], color=color, linewidth=2, + linestyle=':', marker='o', markersize=4) + ax.set_title('Citations to the foundational VIPeR papers', loc='left', + fontsize=17, fontweight='bold', pad=22) + ax.set_ylabel('Citations per year') + ax.set_xticks(range(2010, 2027, 2)) + ax.set_xlim(2009.7, 2026.3) + ax.set_ylim(0, 240) + ax.grid(axis='y', color='#dedede') + ax.set_axisbelow(True) + ax.spines[['top', 'right']].set_visible(False) + ax.legend(frameon=False, loc='upper left') + ax.annotate('213 (2018)', xy=(2018, 213), xytext=(2019, 225), + color='#11637b', fontsize=11, + arrowprops={'arrowstyle': '-', 'color': '#11637b'}) + ax.annotate('179 (2017)', xy=(2017, 179), xytext=(2013, 204), + color='#b64b24', fontsize=11, + arrowprops={'arrowstyle': '-', 'color': '#b64b24'}) + fig.text(0.12, 0.025, + 'Google Scholar · observed October 12, 2026 · 2026 is year-to-date', + fontsize=10, color='#555555') + fig.subplots_adjust(left=0.12, right=0.97, bottom=0.13, top=0.84) + fig.savefig(ASSETS / 'citations-over-time.png') + plt.close(fig) + + +if __name__ == '__main__': + main() diff --git a/build_rank_one_chart.py b/build_rank_one_chart.py new file mode 100644 index 0000000..d524825 --- /dev/null +++ b/build_rank_one_chart.py @@ -0,0 +1,74 @@ +#!/usr/bin/env python3 +"""Render the sourced VIPeR performance scatter plot used in the essay.""" + +import json +from datetime import date +from pathlib import Path + +import matplotlib +matplotlib.use('Agg') +import matplotlib.pyplot as plt + +ASSETS = Path(__file__).resolve().parent / 'site/writing/twenty-years-of-person-re-identification/assets' +STYLES = { + 'historical': ('#11637b', 'o', 'Within-dataset methods'), + 'generalization': ('#19724c', 'o', 'External-data training'), + 'reranked': ('#8a446a', 'D', 'Re-ranking'), + 'audit': ('#727272', 'x', 'Protocol unclear') +} +OFFSETS = { + 'PETS baseline': (0, 28), 'ELF': (-22, 23), 'SDALF': (-25, 28), 'KISSME': (-35, -32), + 'eSDC': (-26, -45), 'SalMatch': (-20, 38), 'SCNCD': (-60, 48), 'LOMO + XQDA': (-18, -80), + 'Metric ensembles': (-80, 40), 'KCCA + DCIA': (-85, 44), 'GOG + XQDA': (0, 40), + 'NFST fusion': (0, 72), 'SNR': (12, -48), 'RaMoE': (4, -35), 'ACL': (-20, 42), + 'ReNorm': (-35, 35), 'CLIP-FGDI': (-85, 38), 'NK3ML': (20, -12), 'RND': (-85, -42), + 'DIMN': (-28, -40), 'DualNorm': (-78, 20), 'OSNet + target fine-tuning': (-6, 42) +} + + +def year_position(row): + if row['date_basis'] == 'year_only': + return row['year'] + posted = date.fromisoformat(row['date']) + start, end = date(posted.year, 1, 1), date(posted.year + 1, 1, 1) + return posted.year + (posted - start).days / (end - start).days + + +def main(): + rows = json.loads((ASSETS / 'rank-one-scatter.json').read_text())['papers'] + assert len({r['method'] for r in rows}) == len(rows) + assert all(0 <= r['rank_one_percent'] <= 100 for r in rows) + plt.rcParams.update({'font.family': 'sans-serif', 'font.size': 12}) + fig, ax = plt.subplots(figsize=(12, 7.2), dpi=180) + fig.patch.set_facecolor('white') + ax.set_facecolor('white') + for setting, (color, marker, label) in STYLES.items(): + for row in (r for r in rows if r['setting'] == setting): + x, rate = year_position(row), row['rank_one_percent'] + shape = '^' if row['method'] == 'NK3ML' else marker + ax.scatter(x, rate, c=color, s=62 if row['highlighted'] else 30, + marker=shape, linewidths=1.2, zorder=3, clip_on=False) + if row['highlighted'] or row['method'] in ('NK3ML', 'RND'): + ax.annotate(row.get('annotation', row['method']), (x, rate), xytext=OFFSETS[row['method']], + textcoords='offset points', fontsize=10, va='center', + ha='right' if row['method'] == 'OSNet + target fine-tuning' else 'left', + arrowprops={'arrowstyle': '-', 'color': color, 'alpha': 0.7, 'linewidth': 0.7}, + bbox={'facecolor': 'white', 'edgecolor': 'none', 'pad': 1, 'alpha': 0.95}, zorder=4) + ax.scatter([], [], c=color, marker=marker, s=40, label=label) + ax.set_xlim(2006.75, 2026.75) + ax.set_xticks(list(range(2007, 2027, 2)) + [2026]) + ax.set_xlabel('Year', labelpad=10) + ax.set_ylim(0, 100) + ax.set_yticks(range(0, 101, 20)) + ax.set_ylabel('Rank-one accuracy (%)', labelpad=10) + ax.grid(color='#ededed', linewidth=0.8) + ax.set_axisbelow(True) + ax.spines[['top', 'right']].set_visible(False) + ax.legend(loc='lower right', fontsize=10, frameon=False) + fig.subplots_adjust(left=0.08, right=0.98, top=0.98, bottom=0.14) + fig.savefig(ASSETS / 'rank-one-scatter.png') + plt.close(fig) + + +if __name__ == '__main__': + main() diff --git a/site/css/site.css b/site/css/site.css new file mode 100644 index 0000000..bc17f03 --- /dev/null +++ b/site/css/site.css @@ -0,0 +1,202 @@ +:root { + color-scheme: light dark; + --bg: #ffffff; + --text: #111111; + --nav-bg: #f0f0f0; + --link: #333333; + --link-hover: #0b5fff; + --button-bg: #ffffff; + --button-text: #111111; + --button-border: #c7c7c7; +} + +/* ── Terminal color theme: ice ───────────────────────────── */ +[data-theme="dark"] { + --term-fg: #b3fff6; + --term-bg: #030e0d; + --term-nav: #020a09; + --term-border: #1a4040; + --term-btn: #0b2020; +} +[data-theme="dark"] { + --bg: var(--term-bg); + --text: var(--term-fg); + --nav-bg: var(--term-nav); + --link: var(--term-fg); + --link-hover: #ffffff; + --button-bg: var(--term-btn); + --button-text: var(--term-fg); + --button-border: var(--term-border); +} + +body { + margin: 0; + font-family: Arial, sans-serif; + display: flex; + background: var(--bg); + color: var(--text); +} +.site-nav { + width: 200px; + position: fixed; + top: 0; bottom: 0; left: 0; + display: flex; + flex-direction: column; + background: var(--nav-bg); + padding: 20px; + box-sizing: border-box; + overflow-y: auto; +} +.site-nav a { + display: block; + margin-bottom: 10px; + text-decoration: none; + color: var(--link); + transition: color 0.2s ease; +} +.site-nav a:hover, nav a:focus { color: var(--link-hover); } +.profile { display: flex; align-items: center; gap: 12px; margin-bottom: 20px; } +.profile-pic { + width: 112px; height: 112px; + border-radius: 50%; + object-fit: cover; + border: 2px solid var(--button-border); + transition: filter 0.3s ease, border-radius 0.3s ease; +} +.nav-divider { border: none; border-top: 1px solid var(--button-border); margin: 12px 0; } +.nav-meta { font-size: 11px; color: var(--link); opacity: 0.6; margin-bottom: 6px; } +.theme-toggle { + width: 100%; + padding: 8px 12px; + margin-top: auto; + border-radius: 6px; + border: 1px solid var(--button-border); + background: var(--button-bg); + color: var(--button-text); + cursor: pointer; + font-size: 14px; + text-align: left; +} +.theme-toggle:focus-visible { outline: 2px solid var(--link-hover); outline-offset: 2px; } +main { margin-left: 220px; padding: 20px; max-width: 800px; } +section { margin-bottom: 60px; } + +/* ── Hobbies poem ─────────────────────────────────────────── */ +.poem { + margin: 24px 0 0; + padding: 4px 0 4px 16px; + border-left: 2px solid var(--button-border); + font-family: Georgia, 'Times New Roman', serif; + font-style: italic; + line-height: 1.55; + color: var(--text); + opacity: 0.85; +} +.poem cite { + display: block; + margin-top: 10px; + font-style: normal; + font-size: 0.9em; + opacity: 0.8; +} +.poem cite a { color: inherit; } +[data-theme="dark"] .poem { + font-family: 'Courier New', Courier, monospace; + font-style: normal; + border-left-color: var(--term-border); +} + +/* ── Terminal dark mode ───────────────────────────────────── */ + +[data-theme="dark"] body { + font-family: 'Courier New', Courier, monospace; + font-size: 14px; + line-height: 1.7; +} + +/* h1: hidden in dark mode — NFO box carries the name */ +[data-theme="dark"] .portfolio h1 { display: none; } + +/* h2: full ASCII box using ::before / ::after with \A newlines */ +[data-theme="dark"] h2 { + font-size: 1em; + font-weight: normal; + text-transform: uppercase; + letter-spacing: 0; + color: inherit; + margin-bottom: 1em; +} +[data-theme="dark"] h2::before { + content: '╔══════════════════════════════════════════════════════╗\A║ '; + white-space: pre; +} +[data-theme="dark"] h2::after { + content: '\A╚══════════════════════════════════════════════════════╝'; + white-space: pre; +} + +[data-theme="dark"] main a { color: inherit; text-decoration: underline; text-underline-offset: 3px; } +[data-theme="dark"] main a:hover { opacity: 0.7; } +[data-theme="dark"] .site-nav a { color: inherit; text-decoration: none; } +[data-theme="dark"] .site-nav a::before { content: '> '; opacity: 0.3; } +[data-theme="dark"] strong { color: inherit; } +[data-theme="dark"] main ul { list-style-type: '* '; padding-left: 1.5em; } +[data-theme="dark"] .profile-pic { + filter: grayscale(100%) brightness(0.65) contrast(1.2); + border-radius: 4px; +} + +/* NFO header — dark mode only */ +.nfo-header { display: none; margin-bottom: 40px; line-height: 1.2; overflow-x: auto; opacity: 0.85; } +[data-theme="dark"] .nfo-header { display: block; } + +/* Terminal splash — dark mode only */ +.terminal-splash { display: none; margin-bottom: 16px; overflow-x: auto; opacity: 0.45; } +[data-theme="dark"] .terminal-splash { display: block; } + + +@media (max-width: 768px) { + body { flex-direction: column; } + .site-nav { width: 100%; position: relative; top: auto; bottom: auto; } + main { margin: 0; padding: 10px; max-width: none; } +} + +.nav-artifacts { + margin-top: 8px; + font-size: 11px; + opacity: 0.6; + display: flex; + gap: 10px; +} +.site-nav a[aria-current="page"] { font-weight: bold; } +[data-theme="dark"] .profile a::before { content: none; } +main { min-width: 0; } +a { overflow-wrap: anywhere; } +[hidden] { display: none !important; } +.subsection-title { margin-top: 32px; } +.page-footer { + margin-top: 48px; + padding-top: 16px; + border-top: 1px solid var(--button-border); + font-size: 13px; +} +@media (max-width: 768px) { + .site-nav { + display: grid; + grid-template-columns: repeat(2, minmax(0, 1fr)); + gap: 0 16px; + padding: 16px; + } + .profile { grid-column: 1 / -1; margin-bottom: 16px; } + .profile-pic { width: 72px; height: 72px; } + .nav-divider { grid-column: 1 / -1; width: 100%; } + .theme-toggle { grid-column: 1 / -1; margin-top: 8px; } + .nav-artifacts { grid-column: 1 / -1; } + main { padding: 16px; } + [data-theme="dark"] h2::before { + content: '+------------------------------+\A| '; + } + [data-theme="dark"] h2::after { + content: '\A+------------------------------+'; + } +} diff --git a/site/css/writing.css b/site/css/writing.css new file mode 100644 index 0000000..bf8ffcd --- /dev/null +++ b/site/css/writing.css @@ -0,0 +1,71 @@ +/* Long-form pages share the portfolio navigation, palette, and theme. */ +.reading-page { box-sizing: border-box; width: calc(100% - 220px); max-width: 860px; padding-top: 36px; } +.reading-page h1 { font-size: 36px; line-height: 1.15; margin: 12px 0 18px; letter-spacing: 0; } +.reading-page h1 span { white-space: nowrap; } +.reading-page h2 { font-size: 24px; line-height: 1.3; margin-top: 42px; letter-spacing: 0; } +.reading-page h3 { font-size: 19px; letter-spacing: 0; } +.breadcrumb { margin: 0 0 22px; font-size: 13px; } +.article-meta, .writing-date { font-size: 13px; opacity: 0.7; } +.article-deck { font-size: 18px; line-height: 1.55; } +.essay { font-size: 17px; line-height: 1.7; } +.essay p { margin: 0 0 22px; } +.essay .lede { font-size: 19px; } +.essay section { margin-bottom: 0; } +.essay sup { line-height: 0; } +.essay sup a { font-size: 12px; padding: 0 2px; text-decoration: none; } +.essay [id], .dataset-page [id] { scroll-margin-top: 24px; } +.reading-page figure { margin: 32px 0; } +.reading-page figure img { display: block; width: 100%; height: auto; } +.reading-page .dataset-figure { width: 100%; max-width: 560px; margin: 32px auto; } +.reading-page figcaption { margin-top: 12px; font-size: 13px; line-height: 1.5; opacity: 0.8; } +.reading-page .chart-figure img { border: 1px solid var(--button-border); background: #fff; } +.milestone-values { margin: 0 0 32px; font-size: 14px; } +.milestone-values summary { cursor: pointer; padding: 8px 0; } +.milestone-values table { margin-top: 12px; } +.milestone-values tbody th { text-align: left; padding-top: 20px; } +.milestone-values td:last-child { white-space: nowrap; } +figure a { display: block; } +.source-notes { border-top: 1px solid var(--button-border); margin-top: 42px; padding-top: 8px; font-size: 14px; } +.source-notes li { margin-bottom: 12px; } +.writing-row { display: grid; grid-template-columns: 150px minmax(0, 1fr); gap: 24px; padding: 24px 0; border-top: 1px solid var(--button-border); } +.writing-row img { display: block; width: 150px; aspect-ratio: 3 / 2; object-fit: cover; } +.writing-row h2 { margin: 6px 0 12px; font-size: 22px; } +.writing-row p { margin: 0 0 8px; line-height: 1.55; } +.writing-row .reading-note { font-size: 13px; opacity: 0.7; } +.dataset-facts { width: 100%; border-collapse: collapse; margin: 24px 0; font-size: 15px; } +.dataset-facts th, .dataset-facts td { text-align: left; vertical-align: top; padding: 10px 12px 10px 0; border-bottom: 1px solid var(--button-border); } +.dataset-facts th { width: 30%; } +.viewpoint-tables { display: grid; grid-template-columns: minmax(0, 1.25fr) minmax(0, 1fr); gap: 28px; margin: 24px 0; align-items: start; } +.viewpoint-counts { width: 100%; border-collapse: collapse; font-size: 14px; font-variant-numeric: tabular-nums; } +.viewpoint-counts caption { text-align: left; font-weight: bold; margin-bottom: 10px; } +.viewpoint-counts th, .viewpoint-counts td { text-align: right; vertical-align: middle; padding: 8px 6px; border-bottom: 1px solid var(--button-border); } +.viewpoint-counts tr > :first-child { text-align: left; padding-left: 0; } +.viewpoint-counts thead th { line-height: 1.4; } +.viewpoint-counts tfoot { font-weight: bold; } +.viewpoint-note { font-size: 13px; } +.dataset-page { line-height: 1.65; } +.dataset-download { display: inline-flex; align-items: center; gap: 8px; font-weight: bold; } +.dataset-download svg { width: 20px; height: 20px; } +.reading-page pre { padding: 16px; border: 1px solid var(--button-border); overflow-x: auto; font-size: 13px; line-height: 1.5; } +.reading-page code { font-family: 'Courier New', Courier, monospace; } +.checksum { font-size: 13px; overflow-wrap: anywhere; word-break: break-word; } +.dataset-outline { display: flex; flex-wrap: wrap; gap: 8px 20px; margin: 24px 0; font-size: 14px; } +[data-theme="dark"] .reading-page h1 { display: block; font-family: inherit; font-size: 30px; } +[data-theme="dark"] .reading-page h2 { padding-bottom: 8px; border-bottom: 1px solid var(--term-border); font-size: 21px; font-weight: bold; text-transform: none; } +[data-theme="dark"] .reading-page h2::before, [data-theme="dark"] .reading-page h2::after { content: none; } +[data-theme="dark"] .essay { font-size: 15px; } +@media (max-width: 768px) { + .reading-page { width: 100%; max-width: none; padding-top: 24px; } + .reading-page h1 { font-size: 30px; } + .reading-page h2 { font-size: 23px; } + .essay { font-size: 16px; } + .essay .lede { font-size: 18px; } + .reading-page .dataset-figure { max-width: none; } + .writing-row { grid-template-columns: 1fr; gap: 16px; } + .writing-row img { width: 100%; max-width: 320px; } + .dataset-facts th { width: 36%; } + .viewpoint-tables { grid-template-columns: minmax(0, 1fr); } +} +@media (max-width: 360px) { + .reading-page h1, [data-theme="dark"] .reading-page h1 { font-size: 26px; } +} diff --git a/site/datasets/viper/VIPeR.v1.0.zip b/site/datasets/viper/VIPeR.v1.0.zip new file mode 100644 index 0000000..6513f1c Binary files /dev/null and b/site/datasets/viper/VIPeR.v1.0.zip differ diff --git a/site/datasets/viper/citation.bib b/site/datasets/viper/citation.bib new file mode 100644 index 0000000..f80de65 --- /dev/null +++ b/site/datasets/viper/citation.bib @@ -0,0 +1,6 @@ +@article{gray2007, + title={Evaluating Appearance Models for Recognition, Reacquisition, and Tracking}, + author={Gray, D. and Brennan, S. and Tao, H.}, + journal={Proc. IEEE International Workshop on Performance Evaluation for Tracking and Surveillance (PETS)}, + year={2007} +} diff --git a/site/datasets/viper/index.html b/site/datasets/viper/index.html new file mode 100644 index 0000000..ff2bc1e --- /dev/null +++ b/site/datasets/viper/index.html @@ -0,0 +1,176 @@ + + + + + + + + VIPeR Dataset - Douglas Gray + + + + + + +
+ +

VIPeR

+

Viewpoint Invariant Pedestrian Recognition

+

We created VIPeR to evaluate appearance models across changes in camera viewpoint, pose, and illumination. It contains two images of each of 632 pedestrians.

+

Download VIPeR v1.0 (ZIP, 18.84 MiB)

+

This is the original archive released on September 22, 2007, alongside our PETS paper. The ZIP is unchanged, byte for byte, including the images, filenames, and original README. See the checksums below to verify your download.

+

Read: Reflections on 20 Years of Person Re-ID

+ + +
+

Dataset details

+ + + + + + + + + + + + + +
TaskAppearance-based person re-identification: rank gallery images for a probe captured from another camera.
Identities632, with two images per identity.
Images1,264 RGB BMP images; 48 pixels wide × 128 pixels high.
StructureVIPeR/cam_a/, VIPeR/cam_b/, and the original README.txt.
PairingMatch filenames by their numeric identity prefix, such as 001 in 001_45.bmp. Identity numbers have gaps.
ViewpointsArbitrary camera viewpoints; most pairs differ by at least 90 degrees. Filenames include coarse viewpoint labels.
CollectionWe collected footage with two cameras in an academic setting. Subjects were notified of the cameras and were not coached.
ReleaseVersion 1.0, September 22, 2007, UCSC Computer Vision Lab.
AuthorsDouglas Gray, Shane Brennan, and Hai Tao.
Archive19,756,977 bytes. Original ZIP and README preserved unchanged.
+
+ Sample paired images showing viewpoint differences in VIPeR +
Twelve identity pairs from the dataset, with one image from each camera in each pair.
+
+

VIPeR is a small, closed-world benchmark with hand-cropped, low-resolution images and only one observation per identity per camera. It does not measure detection performance, long-term tracking, changing clothes, open-world identification, or deployment accuracy. It has no demographic annotations and should not be used to infer demographic fairness.

+

The original archive contains no separate license file. For questions about permissions or the dataset, contact me.

+
+ +
+

Viewpoint distribution

+

We grouped viewpoints into 45-degree bins, combining left/right-symmetric views. Each entry below counts identity pairs, with the larger angle in the row and the smaller angle in the column. The pairs are unordered: these are viewpoint combinations, not camera A/B assignments.

+
+ + + + + + + + + + + + +
Viewpoint combinations
Larger /
smaller
0°45°90°135°
45°16———
90°24147——
135°43724—
180°10353503
Total632 identity pairs
+ + + + + + + + + + +
Viewpoint differences
Angle differencePairs
45°70
90°363
135°96
180°103
Total632
+
+

Original counts from Figures 2 and 3 of our PETS 2007 paper. A dash marks a combination omitted because of symmetry or equal angles. Most pairs (562 of 632) differ by at least 90 degrees.

+

A few filename labels differ from these published bins: identities 000 and 339 have equal angles in both cameras, and 823 has a 315 label. Computing circular angle differences directly from the filenames gives two pairs at 0 degrees and 68 at 45 degrees; the other difference counts are unchanged. Keep these original filenames intact.

+
+ +
+

Evaluation protocol

+

Use 5x2 cross-validation to evaluate appearance matching, keeping both images of each identity in the same fold:

+
    +
  1. Randomly divide the 632 identity pairs into two folds, A and B, of 316 identities each.
  2. +
  3. Train on A and test on B. Then train a fresh model on B and test on A.
  4. +
  5. Repeat with five randomized partitions in total. This gives five rounds, each with two train/test evaluations, or ten evaluations per method.
  6. +
  7. Randomize image-pair order and image mirroring. Within each test fold, use one image per identity as gallery and the other as probe; swap gallery and probe and average both directions.
  8. +
  9. Average cumulative matching characteristic (CMC) curves over the folds and rounds. Report the gallery size alongside the curves.
  10. +
+

Train/test reversal and gallery/probe reversal are separate steps. The first changes which identities the model learns from; the second changes the matching direction within the same test identities.

+

For details, see Section 4.3 of our PETS paper and Section 3.5 of Dietterich (1998), which defines five repetitions of two-fold cross-validation. Five random train/test splits without reversing the folds are a different protocol.

+

There are no fixed train/test split files in the original release. Record random seeds and generated identity assignments when reproducing this protocol. State any different protocol explicitly, including external training data, gallery size, or re-ranking.

+

Use rank one to measure how often the top match is correct. Higher ranks measure whether the correct identity appears in a shortlist for a human operator. For the relationship between these curves, tracking, and reacquisition, see Section 4.2 of our PETS paper.

+
+ +
+

Citation guidelines

+

If you use VIPeR, please cite our dataset and evaluation paper:

+

D. Gray, S. Brennan, and H. Tao, Evaluating Appearance Models for Recognition, Reacquisition, and Tracking, IEEE International Workshop on Performance Evaluation for Tracking and Surveillance (PETS), October 2007.

+
@article{gray2007,
+  title={Evaluating Appearance Models for Recognition,
+         Reacquisition, and Tracking},
+  author={Gray, D. and Brennan, S. and Tao, H.},
+  journal={Proc. IEEE International Workshop on Performance
+           Evaluation for Tracking and Surveillance (PETS)},
+  year={2007}
+}
+

Download BibTeX

+

If discussing or comparing with the ELF method, also cite D. Gray and H. Tao, Viewpoint Invariant Pedestrian Recognition with an Ensemble of Localized Features, ECCV 2008, DOI: 10.1007/978-3-540-88682-2_21.

+
+ +
+

Loading a pair

+

After extracting the ZIP, pair images by identity rather than by directory order. This Python example uses Pillow:

+
from pathlib import Path
+from PIL import Image
+
+root = Path("VIPeR")
+def by_identity(camera):
+  return {
+    path.stem.split("_")[0]: path
+    for path in (root / camera).glob("*.bmp")
+  }
+
+a = by_identity("cam_a")
+b = by_identity("cam_b")
+assert a.keys() == b.keys()
+identity = sorted(a)[0]
+image_a = Image.open(a[identity]).convert("RGB")
+image_b = Image.open(b[identity]).convert("RGB")
+

Treat filename viewpoint labels as coarse annotations, including the label exceptions noted above.

+
+ +
+

Archive checksums

+

SHA-256:

+

857e49928a0c74d2da6fbeef372ea2e42b2bce357f87307d74c6601f1a97f267

+

Historical MD5:

+

1c2d9fc1cc800332567a0da25a1ce68c

+
+ +
+ + + + diff --git a/site/index.html b/site/index.html index 907fa5d..25f8104 100644 --- a/site/index.html +++ b/site/index.html @@ -7,192 +7,10 @@ Douglas Gray - Portfolio - + - - + +
de3ug@github:~$ python build.py --format=web
@@ -212,13 +30,12 @@
-
-
+ diff --git a/site/js/navigation.js b/site/js/navigation.js new file mode 100644 index 0000000..eccec79 --- /dev/null +++ b/site/js/navigation.js @@ -0,0 +1,25 @@ +const siteNavigation = document.querySelector('[data-site-root]'); + +if (siteNavigation) { + const base = siteNavigation.dataset.siteRoot; + const home = base === './'; + const section = (id) => home ? `#${id}` : `${base}#${id}`; + siteNavigation.innerHTML = ` +
+ + Portrait of Douglas Gray + +
+ About + Publications & Service + Experience + Personal, Hobbies & Writing + Contact + + + `; +} diff --git a/site/js/render.js b/site/js/render.js index cb98346..d95f050 100644 --- a/site/js/render.js +++ b/site/js/render.js @@ -58,6 +58,15 @@ function injectJsonLd(r) { 'url': 'https://de3ug.github.io/', 'sameAs': r.social.map(s => s.url) }; + const writing = (r.writing || []).filter(entry => visible(entry, 'web')); + if (writing.length) { + ld.subjectOf = writing.map(entry => ({ + '@type': 'Article', + 'headline': entry.title, + 'url': entry.url, + 'dateCreated': entry.date + })); + } const script = document.createElement('script'); script.type = 'application/ld+json'; script.textContent = JSON.stringify(ld, null, 2); @@ -76,30 +85,30 @@ async function render() { `

${esc(r.about)}

`; // Publications + const mediaHtml = renderList(r.projects); document.querySelector('#publications').innerHTML = - `

Selected Publications and Service

` + + `

Publications and Service

` + + `

Selected Publications

` + `

${renderParts(r.publications_intro.parts)}

` + - renderList(r.publications); - - // Patents - document.querySelector('#patents').innerHTML = - `

Selected Patents

` + + renderList([...r.publications, ...(r.datasets || [])]) + + `

${esc(r.reviewer_service)}

` + + `

Selected Patents

` + `

${esc(r.patents_intro)}

` + - renderList(r.patents); - - // Projects - const projectsHtml = renderList(r.projects); - document.querySelector('#projects').innerHTML = projectsHtml - ? `

Selected Media Mentions

${projectsHtml}` - : ''; + renderList(r.patents) + + (mediaHtml ? `

Selected Media Mentions

${mediaHtml}` : ''); // Experience document.querySelector('#experience').innerHTML = `

Education & Work History

` + renderExperience(r.experience); - // Hobbies - let hobbiesHtml = `

Personal & Hobbies

` + renderList(r.hobbies); + // Personal writing and hobbies + const writingItems = (r.writing || []).map(entry => ({ + ...entry, + parts: [{text: entry.title, url: entry.path || entry.url}] + })); + let hobbiesHtml = `

Personal, Hobbies, and Writing

` + + renderList([...writingItems, ...r.hobbies]); if (r.hobby_poem && visible(r.hobby_poem, 'web')) { const p = r.hobby_poem; const body = p.lines.map(esc).join('
'); diff --git a/site/js/writing.js b/site/js/writing.js new file mode 100644 index 0000000..3051c20 --- /dev/null +++ b/site/js/writing.js @@ -0,0 +1,50 @@ +function writingEsc(value) { + return String(value) + .replace(/&/g, '&') + .replace(//g, '>') + .replace(/"/g, '"'); +} + +function writingVisible(item) { + return !(item.hide_from || []).includes('web'); +} + +async function renderWritingIndex() { + const root = document.querySelector('[data-writing-index]'); + if (!root) return; + + const response = await fetch(root.dataset.resumeSrc); + if (!response.ok) throw new Error(`Could not load writing metadata (${response.status})`); + + const resume = await response.json(); + const entries = (resume.writing || []).filter(writingVisible); + if (!entries.length) { + root.innerHTML = '

No writing has been published yet.

'; + return; + } + + root.innerHTML = entries.map((entry) => { + const href = new URL(entry.path, response.url).href; + const image = new URL(entry.image, response.url).href; + const status = entry.status ? `${entry.status} · ` : ''; + return ` +
+ + ${writingEsc(entry.image_alt)} + +
+

${writingEsc(status + entry.date_display)}

+

${writingEsc(entry.title)}

+

${writingEsc(entry.summary)}

+

${writingEsc(entry.reading_time)} read

+
+
`; + }).join(''); +} + +renderWritingIndex().catch((error) => { + console.error(error); + const root = document.querySelector('[data-writing-index]'); + if (root) root.innerHTML = '

Writing could not be loaded.

'; +}); diff --git a/site/llms.txt b/site/llms.txt index 445489c..14deab0 100644 --- a/site/llms.txt +++ b/site/llms.txt @@ -4,7 +4,11 @@ Research scientist who has led teams and shipped at global scale: the vision-language systems behind Amazon Lens and visual navigation features in Amazon product search reach hundreds of millions of customers. I believe world-model-grounded, full-duplex interfaces are where computing is heading, and want to work on the hardest problems to get there: multimodal reasoning, agentic retrieval, generative models, and real-time interfaces. -Last updated: 09/02/2026 +Last updated: 10/13/2026 + +## Writing + +- [Reflections on 20 Years of Person Re-ID](https://de3ug.github.io/writing/twenty-years-of-person-re-identification/): Why I built VIPeR to measure appearance matching across camera viewpoints, how the field progressed, and the questions that led me away from surveillance research. Contact: - Email: douglas.ryan.gray@gmail.com diff --git a/site/resume.json b/site/resume.json index 50eb295..00caa32 100644 --- a/site/resume.json +++ b/site/resume.json @@ -5,7 +5,21 @@ "_about_extended": "Research scientist who has led teams and shipped vision-language systems at global scale. My work at Amazon has been a series of incremental steps toward a longer-term conviction: that the next great leap in computing isn't a better chat interface, but full-duplex interaction — video, audio, and language flowing through systems that persistently model the world. I think world models are the key technology, and the problems I most want to work on are multimodal reasoning, agentic retrieval, generative models, and real-time interfaces. At Amazon I've led the vision-language systems behind Amazon Lens, which lets customers point a camera at anything and find it, along with visual navigation features in Amazon product search used by hundreds of millions of customers globally. My research has appeared at CVPR, KDD, ECCV, and WACV, and I've co-organized the Multimodal Representation and Retrieval workshop at SIGIR and ICCV.", "location": "Portola Valley, CA (near Palo Alto, 30 min south of SF)", "email": "douglas.ryan.gray@gmail.com", - "last_updated": "09/02/2026", + "last_updated": "10/13/2026", + "writing": [ + { + "title": "Reflections on 20 Years of Person Re-ID", + "url": "https://de3ug.github.io/writing/twenty-years-of-person-re-identification/", + "path": "writing/twenty-years-of-person-re-identification/", + "date": "2026-10-13", + "date_display": "October 2026", + "reading_time": "12 minute", + "summary": "Why I built VIPeR to measure appearance matching across camera viewpoints, how the field progressed, and the questions that led me away from surveillance research.", + "image": "writing/twenty-years-of-person-re-identification/assets/viper-paired-sample.png", + "image_alt": "Paired pedestrian images from the VIPeR dataset showing large changes in viewpoint", + "hide_from": ["resume"] + } + ], "career_chronology": [ { "year": 2026, @@ -64,7 +78,11 @@ ] } ], - "publications_intro": {"parts": ["5,300+ citations \u00b7 h-index\u00a021 (17 since 2019) \u00b7 ", {"text": "Google Scholar", "url": "https://scholar.google.com/citations?user=YliQcmwAAAAJ&hl=en"}, " \u00b7 Reviewer: NeurIPS, CVPR, ICCV, ECCV, WACV"]}, + "publications_intro": {"parts": ["5,400+ citations \u00b7 h-index\u00a021 \u00b7 ", {"text": "Google Scholar", "url": "https://scholar.google.com/citations?user=YliQcmwAAAAJ&hl=en"}]}, + "reviewer_service": "Reviewer for CVPR, ICCV, ECCV, WACV, and NeurIPS.", + "datasets": [ + {"hide_from": ["resume"], "parts": [{"text": "VIPeR dataset", "url": "datasets/viper/"}, " \u00b7 Download, dataset details, and citation guidelines."]} + ], "publications": [ {"parts": [{"text": "Eliciting Self-Verification in Multimodal Reasoning Agents with Reinforcement Learning", "url": "papers/gray_eccv2026_svrl.pdf"}, ", ECCV\u00a02026", {"text": "\u00a0\u00b7\u00a0Supplementary\u00a0material", "url": "papers/gray_eccv2026_svrl_supplementary.pdf", "hide_from": ["resume"]}]}, {"parts": ["Co-organizer, ", {"text": "Multimodal Representation and Retrieval Workshop", "url": "https://mrr-workshop.github.io/"}, " (SIGIR\u00a02024, ICCV 2025)."]}, @@ -74,7 +92,7 @@ {"parts": [{"text": "Hierarchical Proxy-Based Loss for Deep Metric Learning", "url": "https://www.amazon.science/publications/hierarchical-proxy-based-loss-for-deep-metric-learning"}, ", WACV\u00a02022"]}, {"parts": [{"text": "Viewpoint Invariant Pedestrian Recognition with an Ensemble of Localized Features", "url": "papers/gray_eccv2008_pedestrian.pdf"}, ", ECCV\u00a02008 (1,900+ citations)"]}, {"parts": [{"text": "Evaluating Appearance Models for Recognition, Reacquisition, and Tracking", "url": "papers/gray_pets2007_appearance.pdf"}, ", PETS\u00a02007 (1,300+ citations)"]}, - {"hide_from": ["resume", "web"], "parts": ["Reviewer for NeurIPS, CVPR, ICCV, ECCV, and WACV."]} + {"hide_from": ["resume", "web"], "parts": ["Reviewer for CVPR, ICCV, ECCV, WACV, and NeurIPS."]} ], "patents_intro": "Named inventor on 40+ patents in computer vision, visual search, and camera/AR systems.", "patents": [ diff --git a/site/resume.pdf b/site/resume.pdf index 8883ceb..95b12b1 100644 Binary files a/site/resume.pdf and b/site/resume.pdf differ diff --git a/site/writing/index.html b/site/writing/index.html new file mode 100644 index 0000000..c105ac0 --- /dev/null +++ b/site/writing/index.html @@ -0,0 +1,26 @@ + + + + + + + + Writing - Douglas Gray + + + + + +
+ +

Writing

+
+

Loading writing...

+
+ +
+ + + + + diff --git a/site/writing/twenty-years-of-person-re-identification/assets/citations-over-time.png b/site/writing/twenty-years-of-person-re-identification/assets/citations-over-time.png new file mode 100644 index 0000000..47c15b1 Binary files /dev/null and b/site/writing/twenty-years-of-person-re-identification/assets/citations-over-time.png differ diff --git a/site/writing/twenty-years-of-person-re-identification/assets/rank-one-milestones.json b/site/writing/twenty-years-of-person-re-identification/assets/rank-one-milestones.json new file mode 100644 index 0000000..6340693 --- /dev/null +++ b/site/writing/twenty-years-of-person-re-identification/assets/rank-one-milestones.json @@ -0,0 +1,186 @@ +{ + "verified_on": "2026-10-13", + "description": "Selected reported VIPeR rank-one results, not a complete state-of-the-art history or a controlled re-evaluation.", + "gallery_size": 316, + "caveat": "This is a selected history, not a leaderboard. Historical methods and unseen-target domain generalization are separated. Training sources, pretraining, parameter selection, and randomized evaluation procedures differ even within each group. Approximate values are labeled explicitly.", + "results": [ + { + "year": 2007, + "method": "PETS baseline", + "variant": "Hand-localized 3D correlogram, 4 bins", + "rank_one_percent": 5.0, + "approximate": true, + "setting": "historical", + "source": "https://de3ug.github.io/papers/gray_pets2007_appearance.pdf", + "source_location": "Figure 6, right-hand CMC plot, PDF page 5", + "protocol": "Original 5x2 evaluation with 316-image galleries. Approximately 5 percent read from the highest rank-one baseline in the figure; not a tabulated value. This variant includes manually localized body regions." + }, + { + "year": 2008, + "method": "ELF", + "rank_one_percent": 12.0, + "approximate": true, + "setting": "historical", + "source": "https://de3ug.github.io/papers/gray_eccv2008_pedestrian.pdf", + "source_location": "Section 4.2, printed pages 9 and 11; Figure 5", + "protocol": "An even identity train/test split; randomized assignment of paired images to gallery/probe, averaging both directions. The text states around 12 percent." + }, + { + "year": 2010, + "method": "SDALF", + "rank_one_percent": 20.0, + "approximate": true, + "setting": "historical", + "source": "https://iacl.ece.jhu.edu/proceedings/cvpr2010/papers/2173.pdf", + "source_location": "Section 4, VIPeR experiment, PDF page 6; Figure 4", + "protocol": "Average over 10 random sets of 316 pedestrians. Feature weights selected using the first 100 VIPeR pairs. The text states around 20 percent." + }, + { + "year": 2012, + "method": "KISSME", + "rank_one_percent": 19.6, + "approximate": false, + "setting": "historical", + "source": "https://www.tugraz.at/fileadmin/user_upload/Institute/ICG/Documents/lrs/pubs/koestinger_cvpr_2012.pdf", + "source_location": "Section 4.2 and Table 2(a), PDF pages 6-7", + "protocol": "Random 316-identity training and test sets, averaged over 100 runs. The chart uses the average, not the best-run 22 percent result." + }, + { + "year": 2013, + "method": "eSDC", + "variant": "eSDC ocsvm, combined with wHSV and MSCR", + "rank_one_percent": 26.74, + "approximate": false, + "setting": "historical", + "source": "https://openaccess.thecvf.com/content_cvpr_2013/papers/Zhao_Unsupervised_Salience_Learning_2013_CVPR_paper.pdf", + "source_location": "Section 6 and Table 1, PDF pages 6-7", + "protocol": "Half the identities used for fitting without identity labels, half for testing; average of both probe/gallery directions over 10 trials. Uses the best eSDC variant, not the 26.31 percent kNN variant." + }, + { + "year": 2013, + "method": "SalMatch", + "rank_one_percent": 30.16, + "approximate": false, + "setting": "historical", + "source": "https://openaccess.thecvf.com/content_iccv_2013/papers/Zhao_Person_Re-identification_by_2013_ICCV_paper.pdf", + "source_location": "Section 5 and Figure 5(a), PDF pages 6-7", + "protocol": "Disjoint half-identity train/test split, camera A probes and camera B gallery; average over 10 trials. Reported standard deviation 1.23 percentage points." + }, + { + "year": 2014, + "method": "SCNCD", + "variant": "Final(ImgF), color-name and color-histogram fusion", + "rank_one_percent": 37.8, + "approximate": false, + "setting": "historical", + "source": "https://doi.org/10.1007/978-3-319-10590-1_35", + "author_manuscript": "https://www.researchgate.net/publication/278655867_Salient_Color_Names_for_Person_Re-identification", + "source_location": "Sections 4.1-4.2 and Table 1, manuscript pages 9-11", + "protocol": "Half-identity train/test splits, both gallery/probe directions, averaged over 100 trials. The 37.8 percent result includes feature fusion, not standalone SCNCD." + }, + { + "year": 2015, + "method": "LOMO + XQDA", + "rank_one_percent": 40.0, + "approximate": false, + "setting": "historical", + "source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Liao_Person_Re-Identification_by_2015_CVPR_paper.pdf", + "source_location": "Sections 5.1 and 5.1.3, Table 1, PDF pages 5-6", + "protocol": "Random half-identity train/test splits, averaged over 10 trials. Standalone LOMO + XQDA, excluding the separate LADF fusion result." + }, + { + "year": 2015, + "method": "Metric ensembles", + "variant": "Nonlinear CMC top ensemble", + "rank_one_percent": 45.89, + "approximate": false, + "setting": "historical", + "source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Paisitkriangkrai_Learning_to_Rank_2015_CVPR_paper.pdf", + "source_location": "Section 3, Figure 1, PDF pages 6-7", + "protocol": "Half-identity train/test split averaged over 10 partitions; ensemble of multiple features and nonlinear base metrics. Original figure reports 45.89 percent; later tables round to 45.9." + }, + { + "year": 2015, + "method": "KCCA + DCIA", + "rank_one_percent": 63.92, + "approximate": false, + "setting": "reranked", + "source": "https://openaccess.thecvf.com/content_iccv_2015/papers/Garcia_Person_Re-Identification_Ranking_ICCV_2015_paper.pdf", + "source_location": "Sections 4.1-4.3 and Table 2, PDF pages 6-7", + "protocol": "316 training and 316 test identities, 10 random trials. Includes gallery-dependent context analysis and re-ranking; not an independent pairwise matcher." + }, + { + "year": 2016, + "method": "GOG + XQDA", + "variant": "GOGFusion + XQDA, with patch weights and color-space fusion", + "rank_one_percent": 49.7, + "approximate": false, + "setting": "historical", + "source": "https://openaccess.thecvf.com/content_cvpr_2016/papers/Matsukawa_Hierarchical_Gaussian_Descriptor_CVPR_2016_paper.pdf", + "source_location": "Sections 4.1 and 4.3, Table 2, PDF pages 5-7", + "protocol": "Random half-identity train/test splits, averaged over 10 splits. Reported result includes GOG color-space fusion and XQDA metric learning." + }, + { + "year": 2016, + "method": "NFST fusion", + "rank_one_percent": 51.17, + "approximate": false, + "setting": "historical", + "source": "https://openaccess.thecvf.com/content_cvpr_2016/papers/Zhang_Learning_a_Discriminative_CVPR_2016_paper.pdf", + "source_location": "Sections 4.1-4.2 and Table 1, PDF pages 6-7", + "protocol": "Half-identity train/test splits, averaged over 10 trials. Kernelized null-space learning with score-level fusion of LOMO and a second histogram-based descriptor." + }, + { + "year": 2020, + "method": "SNR", + "rank_one_percent": 52.9, + "approximate": false, + "setting": "generalization", + "source": "https://openaccess.thecvf.com/content_CVPR_2020/supplemental/Jin_Style_Normalization_and_CVPR_2020_supplemental.pdf", + "source_location": "Supplementary Table 5, PDF page 5", + "protocol": "Unseen VIPeR target, 10 random evaluation splits; trained on Market-1501, DukeMTMC-reID, CUHK02, CUHK03, and CUHK-SYSU. The alternative four-source setting including MSMT17 reports 55.1 percent (supplementary Table 2); not silently mixed here." + }, + { + "year": 2021, + "method": "RaMoE", + "rank_one_percent": 56.6, + "approximate": false, + "setting": "generalization", + "source": "https://openaccess.thecvf.com/content/CVPR2021/papers/Dai_Generalizable_Person_Re-Identification_With_Relevance-Aware_Mixture_of_Experts_CVPR_2021_paper.pdf", + "source_location": "Section 4.2 and Table 2, PDF pages 6-7", + "protocol": "Protocol 1: unseen VIPeR target, same five source datasets as the SNR point, 10 random evaluation splits. Protocol 2 uses Market-1501, DukeMTMC-reID, CUHK03, and MSMT17 and reports 63.4 percent instead." + }, + { + "year": 2022, + "method": "ACL", + "rank_one_percent": 66.4, + "approximate": false, + "setting": "generalization", + "source": "https://www.ecva.net/papers/eccv_2022/papers_ECCV/papers/136740209.pdf", + "source_location": "Section 4.2 and Table 1, PDF page 10", + "protocol": "Protocol 1: unseen VIPeR target, trained on Market-1501, CUHK02, CUHK03, and CUHK-SYSU; DukeMTMC omitted. Average over 10 random query/gallery splits. Source pool therefore differs from SNR and RaMoE." + }, + { + "year": 2024, + "method": "ReNorm", + "rank_one_percent": 69.1, + "approximate": false, + "setting": "generalization", + "source": "https://www.ecva.net/papers/eccv_2024/papers_ECCV/papers/08753.pdf", + "source_location": "Section 4.1 and Table 1, PDF pages 10-11", + "protocol": "Protocol 1: unseen VIPeR target, trained on Market-1501, CUHK02, CUHK03, and CUHK-SYSU; average of 10 random evaluation splits." + }, + { + "year": 2025, + "method": "CLIP-FGDI", + "rank_one_percent": 76.0, + "approximate": false, + "setting": "generalization", + "source": "https://arxiv.org/abs/2501.16065v3", + "publication": "IEEE TIFS 2025, DOI 10.1109/TIFS.2025.3536608; title spells the acronym CILP-FGDI, while the method/code use CLIP-FGDI", + "source_location": "Section IV-A and Table I, PDF pages 5-6; authors' datasets/viper.py for gallery size", + "implementation_source": "https://github.com/Qi5Lei/CLIP-FGDI/blob/main/datasets/viper.py", + "protocol": "Protocol 1: CLIP-pretrained ViT, trained for re-ID on Market-1501, CUHK02, CUHK03, and CUHK-SYSU, with no VIPeR target training. Authors' loader uses 316-image galleries. Main table reports averages of three experiments; the best hyperparameter ablation in Table VIII reports 77.5 percent, not used as the main milestone." + } + ] +} diff --git a/site/writing/twenty-years-of-person-re-identification/assets/rank-one-over-time.png b/site/writing/twenty-years-of-person-re-identification/assets/rank-one-over-time.png new file mode 100644 index 0000000..66e0bc3 Binary files /dev/null and b/site/writing/twenty-years-of-person-re-identification/assets/rank-one-over-time.png differ diff --git a/site/writing/twenty-years-of-person-re-identification/assets/rank-one-scatter.json b/site/writing/twenty-years-of-person-re-identification/assets/rank-one-scatter.json new file mode 100644 index 0000000..a0d46b5 --- /dev/null +++ b/site/writing/twenty-years-of-person-re-identification/assets/rank-one-scatter.json @@ -0,0 +1,1406 @@ +{ + "observed_on": "2026-10-13", + "description": "Reported VIPeR rank-one results with training protocols and paper references.", + "caveats": [ + "Not exhaustive; secondary comparison values have not all been checked against original papers.", + "Exact arXiv dates can be paper-version dates rather than verified first dates of a secondary cited score.", + "Unknown days remain full-year intervals, not fabricated release dates.", + "Inconsistent/ambiguous claims are retained as audit-only crosses, not accepted performance evidence." + ], + "papers": [ + { + "year": 2007, + "method": "PETS baseline", + "variant": "Hand-localized 3D correlogram, 4 bins", + "rank_one_percent": 5.0, + "approximate": true, + "setting": "historical", + "source": "https://de3ug.github.io/papers/gray_pets2007_appearance.pdf", + "source_location": "Figure 6, right-hand CMC plot, PDF page 5", + "protocol": "Original 5x2 evaluation with 316-image galleries. Approximately 5 percent read from the highest rank-one baseline in the figure; not a tabulated value. This variant includes manually localized body regions.", + "highlighted": true, + "evidence": "original", + "gallery_size": 316, + "date": "2007-07-01", + "date_basis": "year_only", + "date_source": "https://de3ug.github.io/papers/gray_pets2007_appearance.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication." + }, + { + "year": 2008, + "method": "ELF", + "rank_one_percent": 12.0, + "approximate": true, + "setting": "historical", + "source": "https://de3ug.github.io/papers/gray_eccv2008_pedestrian.pdf", + "source_location": "Section 4.2, printed pages 9 and 11; Figure 5", + "protocol": "An even identity train/test split; randomized assignment of paired images to gallery/probe, averaging both directions. The text states around 12 percent.", + "highlighted": true, + "evidence": "original", + "gallery_size": 316, + "date": "2008-07-01", + "date_basis": "year_only", + "date_source": "https://de3ug.github.io/papers/gray_eccv2008_pedestrian.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication." + }, + { + "year": 2010, + "method": "SDALF", + "rank_one_percent": 20.0, + "approximate": true, + "setting": "historical", + "source": "https://iacl.ece.jhu.edu/proceedings/cvpr2010/papers/2173.pdf", + "source_location": "Section 4, VIPeR experiment, PDF page 6; Figure 4", + "protocol": "Average over 10 random sets of 316 pedestrians. Feature weights selected using the first 100 VIPeR pairs. The text states around 20 percent.", + "highlighted": true, + "evidence": "original", + "gallery_size": 316, + "date": "2010-07-01", + "date_basis": "year_only", + "date_source": "https://iacl.ece.jhu.edu/proceedings/cvpr2010/papers/2173.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication." + }, + { + "method": "CPS", + "year": 2011, + "rank_one_percent": 21.84, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://openaccess.thecvf.com/content_cvpr_2013/papers/Zhao_Unsupervised_Salience_Learning_2013_CVPR_paper.pdf", + "source_location": "Table 1, PDF page 7", + "paper_title": "Custom pictorial structures for re-identification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2011-07-01", + "date_basis": "year_only", + "date_source": "https://openaccess.thecvf.com/content_cvpr_2013/papers/Zhao_Unsupervised_Salience_Learning_2013_CVPR_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.5244/c.25.68" + }, + { + "method": "PRDC", + "year": 2011, + "rank_one_percent": 15.66, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://openaccess.thecvf.com/content_cvpr_2013/papers/Zhao_Unsupervised_Salience_Learning_2013_CVPR_paper.pdf", + "source_location": "Table 1, PDF page 7", + "paper_title": "Person re-identification by probabilistic relative distance comparison", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2011-07-01", + "date_basis": "year_only", + "date_source": "https://openaccess.thecvf.com/content_cvpr_2013/papers/Zhao_Unsupervised_Salience_Learning_2013_CVPR_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1109/cvpr.2011.5995598" + }, + { + "year": 2012, + "method": "KISSME", + "rank_one_percent": 19.6, + "approximate": false, + "setting": "historical", + "source": "https://www.tugraz.at/fileadmin/user_upload/Institute/ICG/Documents/lrs/pubs/koestinger_cvpr_2012.pdf", + "source_location": "Section 4.2 and Table 2(a), PDF pages 6-7", + "protocol": "Random 316-identity training and test sets, averaged over 100 runs. The chart uses the average, not the best-run 22 percent result.", + "highlighted": true, + "evidence": "original", + "gallery_size": 316, + "date": "2012-07-01", + "date_basis": "year_only", + "date_source": "https://www.tugraz.at/fileadmin/user_upload/Institute/ICG/Documents/lrs/pubs/koestinger_cvpr_2012.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication." + }, + { + "method": "LDFV", + "year": 2012, + "rank_one_percent": 26.53, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Liao_Person_Re-Identification_by_2015_CVPR_paper.pdf", + "source_location": "Table 1 and fusion footnote, PDF page 6", + "paper_title": "Local descriptors encoded by fisher vectors for person re-identification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2012-07-01", + "date_basis": "year_only", + "date_source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Liao_Person_Re-Identification_by_2015_CVPR_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1007/978-3-642-33863-2_41" + }, + { + "method": "PCCA", + "year": 2012, + "rank_one_percent": 19.27, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Liao_Person_Re-Identification_by_2015_CVPR_paper.pdf", + "source_location": "Table 1 and fusion footnote, PDF page 6", + "paper_title": "Pcca: A new approach for distance learning from sparse pairwise constraints", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2012-07-01", + "date_basis": "year_only", + "date_source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Liao_Person_Re-Identification_by_2015_CVPR_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1109/cvpr.2012.6247987" + }, + { + "method": "RPLM", + "year": 2012, + "rank_one_percent": 27.0, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Liao_Person_Re-Identification_by_2015_CVPR_paper.pdf", + "source_location": "Table 1 and fusion footnote, PDF page 6", + "paper_title": "Relaxed pairwise learned metric for person re-identification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2012-07-01", + "date_basis": "year_only", + "date_source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Liao_Person_Re-Identification_by_2015_CVPR_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1007/978-3-642-33783-3_56" + }, + { + "method": "WELF6 + PRDC", + "year": 2012, + "rank_one_percent": 16.14, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Liao_Person_Re-Identification_by_2015_CVPR_paper.pdf", + "source_location": "Table 1 and fusion footnote, PDF page 6", + "paper_title": "Person re-identification: what features are important?", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2012-07-01", + "date_basis": "year_only", + "date_source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Liao_Person_Re-Identification_by_2015_CVPR_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1007/978-3-642-33863-2_39" + }, + { + "method": "eBiCov", + "year": 2012, + "rank_one_percent": 20.66, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://openaccess.thecvf.com/content_cvpr_2013/papers/Zhao_Unsupervised_Salience_Learning_2013_CVPR_paper.pdf", + "source_location": "Table 1, PDF page 7", + "paper_title": "BiCov: A novel image representation for person re-identification and face verification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2012-07-01", + "date_basis": "year_only", + "date_source": "https://openaccess.thecvf.com/content_cvpr_2013/papers/Zhao_Unsupervised_Salience_Learning_2013_CVPR_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.5244/c.26.57" + }, + { + "method": "ColorInv", + "year": 2013, + "rank_one_percent": 24.21, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Liao_Person_Re-Identification_by_2015_CVPR_paper.pdf", + "source_location": "Table 1 and fusion footnote, PDF page 6", + "paper_title": "Color invariants for person reidentification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2013-07-01", + "date_basis": "year_only", + "date_source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Liao_Person_Re-Identification_by_2015_CVPR_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1109/tpami.2012.246" + }, + { + "method": "LADF", + "year": 2013, + "rank_one_percent": 30.22, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Liao_Person_Re-Identification_by_2015_CVPR_paper.pdf", + "source_location": "Table 1 and fusion footnote, PDF page 6", + "paper_title": "Learning locally-adaptive decision functions for person verification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2013-07-01", + "date_basis": "year_only", + "date_source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Liao_Person_Re-Identification_by_2015_CVPR_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1109/cvpr.2013.463" + }, + { + "method": "LFDA", + "year": 2013, + "rank_one_percent": 24.18, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Liao_Person_Re-Identification_by_2015_CVPR_paper.pdf", + "source_location": "Table 1 and fusion footnote, PDF page 6", + "paper_title": "Local Fisher Discriminant Analysis for Pedestrian Re-identification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2013-07-01", + "date_basis": "year_only", + "date_source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Liao_Person_Re-Identification_by_2015_CVPR_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1109/cvpr.2013.426" + }, + { + "year": 2013, + "method": "SalMatch", + "rank_one_percent": 30.16, + "approximate": false, + "setting": "historical", + "source": "https://openaccess.thecvf.com/content_iccv_2013/papers/Zhao_Person_Re-identification_by_2013_ICCV_paper.pdf", + "source_location": "Section 5 and Figure 5(a), PDF pages 6-7", + "protocol": "Disjoint half-identity train/test split, camera A probes and camera B gallery; average over 10 trials. Reported standard deviation 1.23 percentage points.", + "highlighted": true, + "evidence": "original", + "gallery_size": 316, + "date": "2013-07-01", + "date_basis": "year_only", + "date_source": "https://openaccess.thecvf.com/content_iccv_2013/papers/Zhao_Person_Re-identification_by_2013_ICCV_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication." + }, + { + "year": 2013, + "method": "eSDC", + "variant": "eSDC ocsvm, combined with wHSV and MSCR", + "rank_one_percent": 26.74, + "approximate": false, + "setting": "historical", + "source": "https://openaccess.thecvf.com/content_cvpr_2013/papers/Zhao_Unsupervised_Salience_Learning_2013_CVPR_paper.pdf", + "source_location": "Section 6 and Table 1, PDF pages 6-7", + "protocol": "Half the identities used for fitting without identity labels, half for testing; average of both probe/gallery directions over 10 trials. Uses the best eSDC variant, not the 26.31 percent kNN variant.", + "highlighted": true, + "evidence": "original", + "gallery_size": 316, + "date": "2013-07-01", + "date_basis": "year_only", + "date_source": "https://openaccess.thecvf.com/content_cvpr_2013/papers/Zhao_Unsupervised_Salience_Learning_2013_CVPR_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication." + }, + { + "method": "Mid-level + LADF", + "year": 2014, + "rank_one_percent": 43.39, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Liao_Person_Re-Identification_by_2015_CVPR_paper.pdf", + "source_location": "Table 1 and fusion footnote, PDF page 6", + "paper_title": "Learning mid-level filters for person re-identification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2014-07-01", + "date_basis": "year_only", + "date_source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Liao_Person_Re-Identification_by_2015_CVPR_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1109/cvpr.2014.26", + "alternatives": [ + { + "rank_one_percent": 29.11, + "protocol": "Standalone mid-level representation, without LADF fusion" + } + ] + }, + { + "method": "MtMCML", + "year": 2014, + "rank_one_percent": 28.83, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Liao_Person_Re-Identification_by_2015_CVPR_paper.pdf", + "source_location": "Table 1 and fusion footnote, PDF page 6", + "paper_title": "Person re-identification over camera networks using multi-task distance metric learning", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2014-07-01", + "date_basis": "year_only", + "date_source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Liao_Person_Re-Identification_by_2015_CVPR_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1109/tip.2014.2331755" + }, + { + "year": 2014, + "method": "SCNCD", + "variant": "Final(ImgF), color-name and color-histogram fusion", + "rank_one_percent": 37.8, + "approximate": false, + "setting": "historical", + "source": "https://doi.org/10.1007/978-3-319-10590-1_35", + "author_manuscript": "https://www.researchgate.net/publication/278655867_Salient_Color_Names_for_Person_Re-identification", + "source_location": "Sections 4.1-4.2 and Table 1, manuscript pages 9-11", + "protocol": "Half-identity train/test splits, both gallery/probe directions, averaged over 100 trials. The 37.8 percent result includes feature fusion, not standalone SCNCD.", + "highlighted": true, + "evidence": "original", + "gallery_size": 316, + "date": "2014-07-01", + "date_basis": "year_only", + "date_source": "https://doi.org/10.1007/978-3-319-10590-1_35", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication." + }, + { + "method": "SSCDL", + "year": 2014, + "rank_one_percent": 25.6, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Liao_Person_Re-Identification_by_2015_CVPR_paper.pdf", + "source_location": "Table 1 and fusion footnote, PDF page 6", + "paper_title": "Semi-supervised coupled dictionary learning for person re-identification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2014-07-01", + "date_basis": "year_only", + "date_source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Liao_Person_Re-Identification_by_2015_CVPR_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1109/cvpr.2014.454" + }, + { + "method": "kBiCov", + "year": 2014, + "rank_one_percent": 31.11, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Liao_Person_Re-Identification_by_2015_CVPR_paper.pdf", + "source_location": "Table 1 and fusion footnote, PDF page 6", + "paper_title": "Covariance descriptor based on bio-inspired features for person re-identification and face verification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2014-07-01", + "date_basis": "year_only", + "date_source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Liao_Person_Re-Identification_by_2015_CVPR_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1016/j.imavis.2014.04.002" + }, + { + "method": "kLFDA", + "year": 2014, + "rank_one_percent": 32.3, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "source_location": "Table 4, PDF page 13", + "paper_title": "Person Re-Identification Using Kernel-Based Metric Learning Methods", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2014-07-01", + "date_basis": "year_only", + "date_source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1007/978-3-319-10584-0_1", + "alternatives": [ + { + "rank_one_percent": 22.0, + "protocol": "rPCCA variant from same kernel-metric paper" + } + ] + }, + { + "method": "CSL", + "year": 2015, + "rank_one_percent": 34.8, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "source_location": "Table 4, PDF page 13", + "paper_title": "Person Re-Identification with Correspondence Structure Learning", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2015-04-23", + "date_basis": "arxiv_version", + "date_source": "https://arxiv.org/abs/1504.06243", + "first_posted": "2015-04-23", + "result_version": 1, + "date_note": "Version posting date; secondary-table values are not proven present in that version.", + "paper_reference": "https://doi.org/10.1109/iccv.2015.366" + }, + { + "year": 2015, + "method": "KCCA + DCIA", + "rank_one_percent": 63.92, + "approximate": false, + "setting": "reranked", + "source": "https://openaccess.thecvf.com/content_iccv_2015/papers/Garcia_Person_Re-Identification_Ranking_ICCV_2015_paper.pdf", + "source_location": "Sections 4.1-4.3 and Table 2, PDF pages 6-7", + "protocol": "316 training and 316 test identities, 10 random trials. Includes gallery-dependent context analysis and re-ranking; not an independent pairwise matcher.", + "highlighted": true, + "evidence": "original", + "gallery_size": 316, + "date": "2015-07-01", + "date_basis": "year_only", + "date_source": "https://openaccess.thecvf.com/content_iccv_2015/papers/Garcia_Person_Re-Identification_Ranking_ICCV_2015_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication." + }, + { + "year": 2015, + "method": "LOMO + XQDA", + "rank_one_percent": 40.0, + "approximate": false, + "setting": "historical", + "source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Liao_Person_Re-Identification_by_2015_CVPR_paper.pdf", + "source_location": "Sections 5.1 and 5.1.3, Table 1, PDF pages 5-6", + "protocol": "Random half-identity train/test splits, averaged over 10 trials. Standalone LOMO + XQDA, excluding the separate LADF fusion result.", + "highlighted": true, + "evidence": "original", + "gallery_size": 316, + "date": "2015-05-06", + "date_basis": "arxiv_version", + "date_source": "https://arxiv.org/abs/1406.4216", + "first_posted": "2014-06-17", + "result_version": 2, + "date_note": "Posting date of result-bearing version where verified; otherwise paper version date.", + "earlier_version": { + "date": "2014-06-17", + "rank_one_percent": 34.65, + "method_name": "RQDA", + "source": "https://arxiv.org/abs/1406.4216v1", + "source_location": "Table 2, PDF page 12" + } + }, + { + "method": "MLAPG", + "year": 2015, + "rank_one_percent": 40.7, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "source_location": "Table 4, PDF page 13", + "paper_title": "Efficient PSD Constrained Asymmetric Metric Learning for Person Re-Identification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2015-07-01", + "date_basis": "year_only", + "date_source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1109/iccv.2015.420" + }, + { + "year": 2015, + "method": "Metric ensembles", + "variant": "Nonlinear CMC top ensemble", + "rank_one_percent": 45.89, + "approximate": false, + "setting": "historical", + "source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Paisitkriangkrai_Learning_to_Rank_2015_CVPR_paper.pdf", + "source_location": "Section 3, Figure 1, PDF pages 6-7", + "protocol": "Half-identity train/test split averaged over 10 partitions; ensemble of multiple features and nonlinear base metrics. Original figure reports 45.89 percent; later tables round to 45.9.", + "highlighted": true, + "evidence": "original", + "gallery_size": 316, + "date": "2015-07-01", + "date_basis": "year_only", + "date_source": "https://openaccess.thecvf.com/content_cvpr_2015/papers/Paisitkriangkrai_Learning_to_Rank_2015_CVPR_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication." + }, + { + "method": "PolyMap", + "year": 2015, + "rank_one_percent": 36.8, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "source_location": "Table 4, PDF page 13", + "paper_title": "Similarity learning on an explicit polynomial kernel feature map for person re-identification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2015-07-01", + "date_basis": "year_only", + "date_source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1109/cvpr.2015.7298764" + }, + { + "method": "QALF", + "year": 2015, + "rank_one_percent": 30.2, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "source_location": "Table 4, PDF page 13", + "paper_title": "Query-adaptive late fusion for image search and person re-identification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2015-07-01", + "date_basis": "year_only", + "date_source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1109/cvpr.2015.7298783" + }, + { + "method": "Semantic", + "year": 2015, + "rank_one_percent": 41.6, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "source_location": "Table 4, PDF page 13", + "paper_title": "Transferring a semantic representation for person re-identification and search", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2015-07-01", + "date_basis": "year_only", + "date_source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1109/cvpr.2015.7299046", + "alternatives": [ + { + "rank_one_percent": 31.1, + "protocol": "Different Semantic configuration cited in GOG Table 2" + } + ] + }, + { + "method": "DGD", + "year": 2016, + "rank_one_percent": 38.6, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "source_location": "Table 4, PDF page 13", + "paper_title": "Learning Deep Feature Representations with Domain Guided Dropout for Person Re-identification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2016-04-26", + "date_basis": "arxiv_version", + "date_source": "https://arxiv.org/abs/1604.07528", + "first_posted": "2016-04-26", + "result_version": 1, + "date_note": "Version posting date; secondary-table values are not proven present in that version.", + "paper_reference": "https://doi.org/10.1109/cvpr.2016.140" + }, + { + "method": "Embedding deep metric", + "year": 2016, + "rank_one_percent": 40.9, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "source_location": "Table 4, PDF page 13", + "paper_title": "Embedding deep metric for person re-identification: A study against large variations", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2016-07-01", + "date_basis": "year_only", + "date_source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1007/978-3-319-46448-0_44" + }, + { + "year": 2016, + "method": "GOG + XQDA", + "variant": "GOGFusion + XQDA, with patch weights and color-space fusion", + "rank_one_percent": 49.7, + "approximate": false, + "setting": "historical", + "source": "https://openaccess.thecvf.com/content_cvpr_2016/papers/Matsukawa_Hierarchical_Gaussian_Descriptor_CVPR_2016_paper.pdf", + "source_location": "Sections 4.1 and 4.3, Table 2, PDF pages 5-7", + "protocol": "Random half-identity train/test splits, averaged over 10 splits. Reported result includes GOG color-space fusion and XQDA metric learning.", + "highlighted": true, + "evidence": "original", + "gallery_size": 316, + "date": "2016-07-01", + "date_basis": "year_only", + "date_source": "https://openaccess.thecvf.com/content_cvpr_2016/papers/Matsukawa_Hierarchical_Gaussian_Descriptor_CVPR_2016_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication." + }, + { + "method": "Gated SCNN", + "year": 2016, + "rank_one_percent": 37.8, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "source_location": "Table 4, PDF page 13", + "paper_title": "Gated Siamese Convolutional Neural Network Architecture for Human Re-identification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2016-07-28", + "date_basis": "arxiv_version", + "date_source": "https://arxiv.org/abs/1607.08378", + "first_posted": "2016-07-28", + "result_version": 1, + "date_note": "Version posting date; secondary-table values are not proven present in that version.", + "paper_reference": "https://doi.org/10.1007/978-3-319-46484-8_48" + }, + { + "year": 2016, + "method": "NFST fusion", + "rank_one_percent": 51.17, + "approximate": false, + "setting": "historical", + "source": "https://openaccess.thecvf.com/content_cvpr_2016/papers/Zhang_Learning_a_Discriminative_CVPR_2016_paper.pdf", + "source_location": "Sections 4.1-4.2 and Table 1, PDF pages 6-7", + "protocol": "Half-identity train/test splits, averaged over 10 trials. Kernelized null-space learning with score-level fusion of LOMO and a second histogram-based descriptor.", + "highlighted": true, + "evidence": "original", + "gallery_size": 316, + "date": "2016-03-07", + "date_basis": "arxiv_version", + "date_source": "https://arxiv.org/abs/1603.02139", + "first_posted": "2016-03-07", + "result_version": 1, + "date_note": "Posting date of result-bearing version where verified; otherwise paper version date." + }, + { + "method": "S-LSTM", + "year": 2016, + "rank_one_percent": 42.4, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "source_location": "Table 4, PDF page 13", + "paper_title": "A Siamese Long Short-Term Memory Architecture for Human Re-identification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2016-07-28", + "date_basis": "arxiv_version", + "date_source": "https://arxiv.org/abs/1607.08381", + "first_posted": "2016-07-28", + "result_version": 1, + "date_note": "Version posting date; secondary-table values are not proven present in that version.", + "paper_reference": "https://doi.org/10.1007/978-3-319-46478-7_9" + }, + { + "method": "SCSP", + "year": 2016, + "rank_one_percent": 53.5, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "source_location": "Table 4, PDF page 13", + "paper_title": "Similarity Learning with Spatial Constraints for Person Re-identification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2016-07-01", + "date_basis": "year_only", + "date_source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1109/cvpr.2016.142" + }, + { + "method": "SSDAL", + "year": 2016, + "rank_one_percent": 43.5, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "source_location": "Table 4, PDF page 13", + "paper_title": "Deep attributes driven multi-camera person re-identification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2016-07-01", + "date_basis": "year_only", + "date_source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication." + }, + { + "method": "SSSVM", + "year": 2016, + "rank_one_percent": 42.1, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "source_location": "Table 4, PDF page 13", + "paper_title": "Sample-Specific SVM Learning for Person Re-identification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2016-07-01", + "date_basis": "year_only", + "date_source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1109/cvpr.2016.143" + }, + { + "method": "TMA", + "year": 2016, + "rank_one_percent": 48.2, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "source_location": "Table 4, PDF page 13", + "paper_title": "Temporal model adaptation for person reidentification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2016-07-01", + "date_basis": "year_only", + "date_source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication." + }, + { + "method": "TPC", + "year": 2016, + "rank_one_percent": 47.8, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "source_location": "Table 4, PDF page 13", + "paper_title": "Person Re-identification by Multi-Channel Parts-Based CNN with Improved Triplet Loss Function", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2016-07-01", + "date_basis": "year_only", + "date_source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1109/cvpr.2016.149" + }, + { + "method": "l1-graph", + "year": 2016, + "rank_one_percent": 41.5, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "source_location": "Table 4, PDF page 13", + "paper_title": "Person re-identification by unsupervised l1 graph learning", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2016-07-01", + "date_basis": "year_only", + "date_source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication." + }, + { + "method": "CAMEL", + "year": 2017, + "rank_one_percent": 30.9, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "source_location": "Table 4, PDF page 13", + "paper_title": "Cross-View Asymmetric Metric Learning for Unsupervised Person Re-Identification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2017-08-27", + "date_basis": "arxiv_version", + "date_source": "https://arxiv.org/abs/1708.08062", + "first_posted": "2017-08-27", + "result_version": 1, + "date_note": "Version posting date; secondary-table values are not proven present in that version.", + "paper_reference": "https://doi.org/10.1109/iccv.2017.113" + }, + { + "method": "DLPAR", + "year": 2017, + "rank_one_percent": 48.7, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "source_location": "Table 4, PDF page 13", + "paper_title": "Deeply-Learned Part-Aligned Representations for Person Re-Identification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2017-07-23", + "date_basis": "arxiv_version", + "date_source": "https://arxiv.org/abs/1707.07256", + "first_posted": "2017-07-23", + "result_version": 1, + "date_note": "Version posting date; secondary-table values are not proven present in that version.", + "paper_reference": "https://doi.org/10.1109/iccv.2017.349" + }, + { + "method": "MuDeep", + "year": 2017, + "rank_one_percent": 43.0, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "source_location": "Table 4, PDF page 13", + "paper_title": "MuDeep, ICCV 2017, as referenced in NK3ML Table 4", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2017-07-01", + "date_basis": "year_only", + "date_source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication." + }, + { + "method": "OL-MANS", + "year": 2017, + "rank_one_percent": 45.0, + "setting": "reranked", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "source_location": "Table 4, PDF page 13", + "paper_title": "Efficient Online Local Metric Adaptation via Negative Samples for Person Re-identification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2017-07-01", + "date_basis": "year_only", + "date_source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1109/iccv.2017.265" + }, + { + "method": "PDC", + "year": 2017, + "rank_one_percent": 51.3, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "source_location": "Table 4, PDF page 13", + "paper_title": "Pose-Driven Deep Convolutional Model for Person Re-identification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2017-09-25", + "date_basis": "arxiv_version", + "date_source": "https://arxiv.org/abs/1709.08325", + "first_posted": "2017-09-25", + "result_version": 1, + "date_note": "Version posting date; secondary-table values are not proven present in that version.", + "paper_reference": "https://doi.org/10.1109/iccv.2017.427" + }, + { + "method": "SHaPE", + "year": 2017, + "rank_one_percent": 62.0, + "setting": "reranked", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "source_location": "Table 4, PDF page 13", + "paper_title": "SHaPE: A Novel Graph Theoretic Algorithm for Making Consensus-Based Decisions in Person Re-identification Systems", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2017-07-01", + "date_basis": "year_only", + "date_source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1109/iccv.2017.127" + }, + { + "method": "SSM", + "year": 2017, + "rank_one_percent": 53.7, + "setting": "reranked", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "source_location": "Table 4, PDF page 13", + "paper_title": "Scalable Person Re-identification on Supervised Smoothed Manifold", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2017-03-24", + "date_basis": "arxiv_version", + "date_source": "https://arxiv.org/abs/1703.08359", + "first_posted": "2017-03-24", + "result_version": 1, + "date_note": "Version posting date; secondary-table values are not proven present in that version.", + "paper_reference": "https://doi.org/10.1109/cvpr.2017.358" + }, + { + "method": "SpindleNet", + "year": 2017, + "rank_one_percent": 53.8, + "setting": "historical", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "source_location": "Table 4, PDF page 13", + "paper_title": "Spindle Net: Person Re-identification with Human Body Region Guided Feature Decomposition and Fusion", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2017-07-01", + "date_basis": "year_only", + "date_source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1109/cvpr.2017.103" + }, + { + "method": "NK3ML", + "year": 2018, + "rank_one_percent": 99.8, + "setting": "historical", + "highlighted": false, + "evidence": "original", + "source": "https://www.ecva.net/papers/eccv_2018/papers_ECCV/papers/T_M_Feroz_Ali_Maximum_Margin_Metric_ECCV_2018_paper.pdf", + "source_location": "Table 4, PDF page 13", + "gallery_size": 316, + "protocol": "Conventional disjoint half-identity split claimed by the original paper; not independently reproduced here.", + "annotation": "NK3ML 99.8%\nreported claim", + "date": "2018-07-28", + "date_basis": "arxiv_version", + "date_source": "https://arxiv.org/abs/1807.10908", + "first_posted": "2018-07-28", + "result_version": 1, + "date_note": "Posting date of result-bearing version where verified; otherwise paper version date." + }, + { + "method": "DIMN", + "year": 2019, + "rank_one_percent": 51.23, + "setting": "generalization", + "highlighted": true, + "evidence": "original", + "source": "https://openaccess.thecvf.com/content_CVPR_2019/papers/Song_Generalizable_Person_Re-Identification_by_Domain-Invariant_Mapping_Network_CVPR_2019_paper.pdf", + "source_location": "Section 4.1 and Table 2, PDF pages 5-6", + "gallery_size": 316, + "paper_title": "Generalizable Person Re-Identification by Domain-Invariant Mapping Network", + "protocol": "Unseen VIPeR target; trained on five external re-ID datasets. Five random half-identity splits plus their swapped versions. Original Table 2 reports 51.23 percent.", + "date": "2019-07-01", + "date_basis": "year_only", + "date_source": "https://openaccess.thecvf.com/content_CVPR_2019/papers/Song_Generalizable_Person_Re-Identification_by_Domain-Invariant_Mapping_Network_CVPR_2019_paper.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1109/cvpr.2019.00081" + }, + { + "method": "DualNorm", + "year": 2019, + "rank_one_percent": 59.4, + "setting": "generalization", + "highlighted": true, + "evidence": "original", + "source": "https://arxiv.org/abs/1905.03422v3", + "source_location": "Section 4.1 and Table 5, PDF pages 5 and 10", + "gallery_size": 316, + "paper_title": "Frustratingly Easy Person Re-Identification: Generalizing Person Re-ID in Practice", + "protocol": "Unseen VIPeR target; externally trained, ImageNet-pretrained ResNet50 + DualNorm configuration, 10 random 316-image gallery splits. The MobileNetV2 main configuration reports 53.9 percent. Earlier v1 ResNet50 result was 54.3 percent; 59.4 is verified in v3, not assigned to the first posting.", + "date": "2019-07-22", + "date_basis": "arxiv_version", + "date_source": "https://arxiv.org/abs/1905.03422", + "first_posted": "2019-05-09", + "result_version": 3, + "date_note": "Posting date of result-bearing version where verified; otherwise paper version date.", + "paper_reference": "https://doi.org/10.48550/arxiv.1905.03422", + "alternatives": [ + { + "rank_one_percent": 53.9, + "protocol": "MobileNetV2 main configuration, v3 Table 1" + }, + { + "rank_one_percent": 54.3, + "date": "2019-05-09", + "protocol": "ResNet50 earlier v1 Table 5, different training configuration" + } + ] + }, + { + "method": "Metric BoW + Null Space", + "year": 2019, + "rank_one_percent": 50.0, + "setting": "historical", + "highlighted": false, + "evidence": "original", + "source": "https://www.scitepress.org/PublishedPapers/2019/72511/pdf/index.html", + "source_location": "Section 4.1.1 and Table 1, printed pages 302-303", + "gallery_size": 316, + "paper_title": "Metric Learning in Codebook Generation of Bag-of-Words for Person Re-identification", + "protocol": "Supervised codebook and null-space metric learning; random half-identity train/test partitions, opposite-camera matching, 10 trials. Original Table 1 reports 50.0 percent.", + "date": "2019-07-01", + "date_basis": "year_only", + "date_source": "https://www.scitepress.org/PublishedPapers/2019/72511/pdf/index.html", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication." + }, + { + "method": "OSNet + target fine-tuning", + "year": 2019, + "rank_one_percent": 68.0, + "setting": "historical", + "highlighted": true, + "evidence": "original", + "source": "https://arxiv.org/abs/1910.06827v1", + "source_location": "Section 4.1 and Table 4, PDF pages 7-8", + "gallery_size": 316, + "paper_title": "Learning Generalisable Omni-Scale Representations for Person Re-Identification", + "annotation": "OSNet\nVIPeR fine-tuning", + "protocol": "2019 first preprint of the expanded OSNet study. External re-ID pretraining followed by VIPeR target fine-tuning, 10 random splits. Not an unseen-target result; not attributed to the earlier ICCV paper without its own VIPeR evidence.", + "date": "2019-10-15", + "date_basis": "arxiv_version", + "date_source": "https://arxiv.org/abs/1910.06827", + "first_posted": "2019-10-15", + "result_version": 1, + "date_note": "Posting date of result-bearing version where verified; otherwise paper version date.", + "paper_reference": "https://doi.org/10.48550/arxiv.1910.06827" + }, + { + "method": "SE + DWE TriNet", + "year": 2019, + "rank_one_percent": 44.9, + "setting": "audit", + "highlighted": false, + "evidence": "original", + "source": "https://hubertshum.com/publications/wscg2019reidentification/files/wscg2019reidentification.pdf", + "source_location": "Section 4.1 and Tables 1 and 4, PDF pages 5-7", + "gallery_size": null, + "paper_title": "Triplet Loss with Channel Attention for Person Re-identification", + "protocol": "Authors train the VIPeR models themselves and report 44.9 percent. The relevant experiment description does not specify the identity split, gallery size, or number of trials; retained as a protocol-audit point, not confirmed comparable evidence.", + "date": "2019-07-01", + "date_basis": "year_only", + "date_source": "https://hubertshum.com/publications/wscg2019reidentification/files/wscg2019reidentification.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.24132/jwscg.2019.27.2.9" + }, + { + "year": 2020, + "method": "SNR", + "rank_one_percent": 52.9, + "approximate": false, + "setting": "generalization", + "source": "https://openaccess.thecvf.com/content_CVPR_2020/supplemental/Jin_Style_Normalization_and_CVPR_2020_supplemental.pdf", + "source_location": "Supplementary Table 5, PDF page 5", + "protocol": "Unseen VIPeR target, 10 random evaluation splits; trained on Market-1501, DukeMTMC-reID, CUHK02, CUHK03, and CUHK-SYSU. The alternative four-source setting including MSMT17 reports 55.1 percent (supplementary Table 2); not silently mixed here.", + "highlighted": true, + "evidence": "original", + "gallery_size": 316, + "date": "2020-05-22", + "date_basis": "arxiv_version", + "date_source": "https://arxiv.org/abs/2005.11037", + "first_posted": "2020-05-22", + "result_version": 1, + "date_note": "Posting date of result-bearing version where verified; otherwise paper version date.", + "alternatives": [ + { + "rank_one_percent": 55.1, + "protocol": "Four-source pool including MSMT17, supplementary Table 2" + } + ] + }, + { + "method": "DMG-Net", + "year": 2021, + "rank_one_percent": 53.9, + "setting": "generalization", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2024/papers_ECCV/papers/08753.pdf", + "source_location": "Table 1, PDF page 11", + "paper_title": "Dual-meta generalization network for generalizable person re-identification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2021-07-01", + "date_basis": "year_only", + "date_source": "https://www.ecva.net/papers/eccv_2024/papers_ECCV/papers/08753.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication." + }, + { + "method": "M3L", + "year": 2021, + "rank_one_percent": 60.8, + "setting": "generalization", + "highlighted": false, + "evidence": "comparison", + "source": "https://arxiv.org/abs/2501.16065v3", + "source_location": "Table I, PDF page 6", + "paper_title": "Learning to Generalize Unseen Domains via Memory-based Multi-Source Meta-Learning for Person Re-Identification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2021-07-01", + "date_basis": "year_only", + "date_source": "https://arxiv.org/abs/2501.16065v3", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication." + }, + { + "method": "MetaBIN", + "year": 2021, + "rank_one_percent": 55.9, + "setting": "generalization", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2024/papers_ECCV/papers/08753.pdf", + "source_location": "Table 1, PDF page 11", + "paper_title": "Meta Batch-Instance Normalization for Generalizable Person Re-Identification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2020-11-30", + "date_basis": "arxiv_version", + "date_source": "https://arxiv.org/abs/2011.14670", + "first_posted": "2020-11-30", + "result_version": 1, + "date_note": "Version posting date; secondary-table values are not proven present in that version.", + "paper_reference": "https://doi.org/10.1109/cvpr46437.2021.00343" + }, + { + "year": 2021, + "method": "RaMoE", + "rank_one_percent": 56.6, + "approximate": false, + "setting": "generalization", + "source": "https://openaccess.thecvf.com/content/CVPR2021/papers/Dai_Generalizable_Person_Re-Identification_With_Relevance-Aware_Mixture_of_Experts_CVPR_2021_paper.pdf", + "source_location": "Section 4.2 and Table 2, PDF pages 6-7", + "protocol": "Protocol 1: unseen VIPeR target, same five source datasets as the SNR point, 10 random evaluation splits. Protocol 2 uses Market-1501, DukeMTMC-reID, CUHK03, and MSMT17 and reports 63.4 percent instead.", + "highlighted": true, + "evidence": "original", + "gallery_size": 316, + "date": "2021-05-19", + "date_basis": "arxiv_version", + "date_source": "https://arxiv.org/abs/2105.09156", + "first_posted": "2021-05-19", + "result_version": 1, + "date_note": "Posting date of result-bearing version where verified; otherwise paper version date.", + "alternatives": [ + { + "rank_one_percent": 63.4, + "protocol": "Protocol 2, different source pool, Table 2" + } + ] + }, + { + "year": 2022, + "method": "ACL", + "rank_one_percent": 66.4, + "approximate": false, + "setting": "generalization", + "source": "https://www.ecva.net/papers/eccv_2022/papers_ECCV/papers/136740209.pdf", + "source_location": "Section 4.2 and Table 1, PDF page 10", + "protocol": "Protocol 1: unseen VIPeR target, trained on Market-1501, CUHK02, CUHK03, and CUHK-SYSU; DukeMTMC omitted. Average over 10 random query/gallery splits. Source pool therefore differs from SNR and RaMoE.", + "highlighted": true, + "evidence": "original", + "gallery_size": 316, + "date": "2022-07-01", + "date_basis": "year_only", + "date_source": "https://www.ecva.net/papers/eccv_2022/papers_ECCV/papers/136740209.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication." + }, + { + "method": "DTIN-Net", + "year": 2022, + "rank_one_percent": 62.9, + "setting": "generalization", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2024/papers_ECCV/papers/08753.pdf", + "source_location": "Table 1, PDF page 11", + "paper_title": "DTIN-Net, ECCV 2022, as referenced in ReNorm Table 1", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2022-07-01", + "date_basis": "year_only", + "date_source": "https://www.ecva.net/papers/eccv_2024/papers_ECCV/papers/08753.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication." + }, + { + "method": "META", + "year": 2022, + "rank_one_percent": 61.5, + "setting": "generalization", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2024/papers_ECCV/papers/08753.pdf", + "source_location": "Table 1, PDF page 11", + "paper_title": "Mimic Embedding via Adaptive Aggregation: Learning Generalizable Person Re-identification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2022-07-01", + "date_basis": "year_only", + "date_source": "https://www.ecva.net/papers/eccv_2024/papers_ECCV/papers/08753.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1007/978-3-031-19781-9_22" + }, + { + "method": "CINorm", + "year": 2023, + "rank_one_percent": 67.6, + "setting": "generalization", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2024/papers_ECCV/papers/08753.pdf", + "source_location": "Table 1, PDF page 11", + "paper_title": "CINorm, TMM 2023, as referenced in ReNorm Table 1", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2023-07-01", + "date_basis": "year_only", + "date_source": "https://www.ecva.net/papers/eccv_2024/papers_ECCV/papers/08753.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication." + }, + { + "method": "DSM + SHS", + "year": 2023, + "rank_one_percent": 62.8, + "setting": "generalization", + "highlighted": false, + "evidence": "comparison", + "source": "https://arxiv.org/abs/2501.16065v3", + "source_location": "Table I, PDF page 6", + "paper_title": "Style-controllable generalized person re-identification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2023-07-01", + "date_basis": "year_only", + "date_source": "https://arxiv.org/abs/2501.16065v3", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication." + }, + { + "method": "GMN", + "year": 2024, + "rank_one_percent": 69.0, + "setting": "generalization", + "highlighted": false, + "evidence": "comparison", + "source": "https://arxiv.org/abs/2501.16065v3", + "source_location": "Table I, PDF page 6", + "paper_title": "Generalizable Metric Network for Cross-Domain Person Re-Identification", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2024-07-01", + "date_basis": "year_only", + "date_source": "https://arxiv.org/abs/2501.16065v3", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1109/tcsvt.2024.3395411" + }, + { + "method": "PAOA", + "year": 2024, + "rank_one_percent": 66.7, + "setting": "generalization", + "highlighted": false, + "evidence": "comparison", + "source": "https://www.ecva.net/papers/eccv_2024/papers_ECCV/papers/08753.pdf", + "source_location": "Table 1, PDF page 11", + "paper_title": "Mitigate Domain Shift by Primary-Auxiliary Objectives Association for Generalizing Person ReID", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2024-07-01", + "date_basis": "year_only", + "date_source": "https://www.ecva.net/papers/eccv_2024/papers_ECCV/papers/08753.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1109/wacv57701.2024.00046" + }, + { + "method": "ReFID", + "year": 2024, + "rank_one_percent": 60.9, + "setting": "generalization", + "highlighted": false, + "evidence": "comparison", + "source": "https://arxiv.org/abs/2501.16065v3", + "source_location": "Table I, PDF page 6", + "paper_title": "ReFID: Reciprocal Frequency-Aware Generalizable Person Re-Identification via Decomposition and Filtering", + "gallery_size": 316, + "protocol": "Cited comparison configuration; original experiment and exact training provenance not yet re-audited.", + "date": "2024-07-01", + "date_basis": "year_only", + "date_source": "https://arxiv.org/abs/2501.16065v3", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication.", + "paper_reference": "https://doi.org/10.1145/3643684" + }, + { + "year": 2024, + "method": "ReNorm", + "rank_one_percent": 69.1, + "approximate": false, + "setting": "generalization", + "source": "https://www.ecva.net/papers/eccv_2024/papers_ECCV/papers/08753.pdf", + "source_location": "Section 4.1 and Table 1, PDF pages 10-11", + "protocol": "Protocol 1: unseen VIPeR target, trained on Market-1501, CUHK02, CUHK03, and CUHK-SYSU; average of 10 random evaluation splits.", + "highlighted": true, + "evidence": "original", + "gallery_size": 316, + "date": "2024-07-01", + "date_basis": "year_only", + "date_source": "https://www.ecva.net/papers/eccv_2024/papers_ECCV/papers/08753.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication." + }, + { + "method": "Attention + graph", + "year": 2025, + "rank_one_percent": 67.2, + "setting": "audit", + "highlighted": false, + "evidence": "original", + "source": "https://publisher.resbee.org/mr/archive/v8i2/a3/p3.pdf", + "source_location": "Table 3 and CMC table", + "gallery_size": null, + "protocol": "Claim only: unclear split/gallery; reported CMC rank-5 exceeds rank-10. Not accepted as benchmark evidence.", + "date": "2025-07-01", + "date_basis": "year_only", + "date_source": "https://publisher.resbee.org/mr/archive/v8i2/a3/p3.pdf", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication." + }, + { + "year": 2025, + "method": "CLIP-FGDI", + "rank_one_percent": 76.0, + "approximate": false, + "setting": "generalization", + "source": "https://arxiv.org/abs/2501.16065v3", + "publication": "IEEE TIFS 2025, DOI 10.1109/TIFS.2025.3536608; title spells the acronym CILP-FGDI, while the method/code use CLIP-FGDI", + "source_location": "Section IV-A and Table I, PDF pages 5-6; authors' datasets/viper.py for gallery size", + "implementation_source": "https://github.com/Qi5Lei/CLIP-FGDI/blob/main/datasets/viper.py", + "protocol": "Protocol 1: CLIP-pretrained ViT, trained for re-ID on Market-1501, CUHK02, CUHK03, and CUHK-SYSU, with no VIPeR target training. Authors' loader uses 316-image galleries. Main table reports averages of three experiments; the best hyperparameter ablation in Table VIII reports 77.5 percent, not used as the main milestone.", + "highlighted": true, + "evidence": "original", + "gallery_size": 316, + "date": "2025-01-27", + "date_basis": "arxiv_version", + "date_source": "https://arxiv.org/abs/2501.16065", + "first_posted": "2025-01-27", + "result_version": 1, + "date_note": "Posting date of result-bearing version where verified; otherwise paper version date.", + "alternatives": [ + { + "rank_one_percent": 77.5, + "protocol": "Best hyperparameter ablation, Table VIII; not main comparison" + } + ] + }, + { + "method": "DictRWMB", + "year": 2025, + "rank_one_percent": 56.9, + "setting": "historical", + "highlighted": false, + "evidence": "original", + "source": "https://www.scitepress.org/publishedPapers/2025/133199/pdf/index.html", + "source_location": "Table 1 (p=316)", + "gallery_size": 316, + "protocol": "Dictionary learning / polynomial ranking, reported p=316; full protocol still to re-audit.", + "date": "2025-07-01", + "date_basis": "year_only", + "date_source": "https://www.scitepress.org/publishedPapers/2025/133199/pdf/index.html", + "date_note": "Year known; midpoint used only to draw a full-year uncertainty interval. Not an asserted July publication." + }, + { + "method": "RND", + "year": 2026, + "rank_one_percent": 73.8, + "setting": "audit", + "highlighted": false, + "evidence": "original", + "source": "https://www.nature.com/articles/s41598-025-30664-4", + "source_location": "Tables 1, 3, and 5", + "gallery_size": 316, + "date": "2026-01-22", + "date_basis": "publisher_publication", + "date_source": "https://www.nature.com/articles/s41598-025-30664-4", + "protocol": "VIPeR training/validation images listed, yet same result labeled unseen-target DG. Correspondence between experiments unclear." + } + ] +} diff --git a/site/writing/twenty-years-of-person-re-identification/assets/rank-one-scatter.png b/site/writing/twenty-years-of-person-re-identification/assets/rank-one-scatter.png new file mode 100644 index 0000000..6e20315 Binary files /dev/null and b/site/writing/twenty-years-of-person-re-identification/assets/rank-one-scatter.png differ diff --git a/site/writing/twenty-years-of-person-re-identification/assets/scholar-citations.json b/site/writing/twenty-years-of-person-re-identification/assets/scholar-citations.json new file mode 100644 index 0000000..292c83c --- /dev/null +++ b/site/writing/twenty-years-of-person-re-identification/assets/scholar-citations.json @@ -0,0 +1,108 @@ +{ + "source": "Google Scholar article citation histograms", + "observed_on": "2026-10-12", + "plotted_range": [ + 2010, + 2026 + ], + "notes": "Separate annual series as displayed by Scholar; not a deduplicated count of papers citing either work. 2026 is year-to-date. Earlier years are outside the common displayed range.", + "papers": { + "eccv2008": { + "title": "Viewpoint Invariant Pedestrian Recognition with an Ensemble of Localized Features", + "total_citations": 2016, + "url": "https://scholar.google.com/citations?view_op=view_citation&hl=en&user=YliQcmwAAAAJ&citation_for_view=YliQcmwAAAAJ:u5HHmVD_uO8C" + }, + "pets2007": { + "title": "Evaluating Appearance Models for Recognition, Reacquisition, and Tracking", + "total_citations": 1405, + "url": "https://scholar.google.com/citations?view_op=view_citation&hl=en&user=YliQcmwAAAAJ&citation_for_view=YliQcmwAAAAJ:UeHWp8X0CEIC" + } + }, + "annual": [ + { + "year": 2010, + "eccv2008": 22, + "pets2007": 18 + }, + { + "year": 2011, + "eccv2008": 49, + "pets2007": 27 + }, + { + "year": 2012, + "eccv2008": 73, + "pets2007": 62 + }, + { + "year": 2013, + "eccv2008": 75, + "pets2007": 58 + }, + { + "year": 2014, + "eccv2008": 141, + "pets2007": 109 + }, + { + "year": 2015, + "eccv2008": 140, + "pets2007": 121 + }, + { + "year": 2016, + "eccv2008": 141, + "pets2007": 147 + }, + { + "year": 2017, + "eccv2008": 200, + "pets2007": 179 + }, + { + "year": 2018, + "eccv2008": 213, + "pets2007": 145 + }, + { + "year": 2019, + "eccv2008": 200, + "pets2007": 120 + }, + { + "year": 2020, + "eccv2008": 169, + "pets2007": 106 + }, + { + "year": 2021, + "eccv2008": 130, + "pets2007": 75 + }, + { + "year": 2022, + "eccv2008": 129, + "pets2007": 52 + }, + { + "year": 2023, + "eccv2008": 90, + "pets2007": 50 + }, + { + "year": 2024, + "eccv2008": 91, + "pets2007": 57 + }, + { + "year": 2025, + "eccv2008": 92, + "pets2007": 43 + }, + { + "year": 2026, + "eccv2008": 50, + "pets2007": 24 + } + ] +} diff --git a/site/writing/twenty-years-of-person-re-identification/assets/viper-paired-sample.png b/site/writing/twenty-years-of-person-re-identification/assets/viper-paired-sample.png new file mode 100644 index 0000000..68d056c Binary files /dev/null and b/site/writing/twenty-years-of-person-re-identification/assets/viper-paired-sample.png differ diff --git a/site/writing/twenty-years-of-person-re-identification/index.html b/site/writing/twenty-years-of-person-re-identification/index.html new file mode 100644 index 0000000..525bf12 --- /dev/null +++ b/site/writing/twenty-years-of-person-re-identification/index.html @@ -0,0 +1,234 @@ + + + + + + + + + + + + + Reflections on 20 Years of Person Re-ID - Douglas Gray + + + + + + +
+ +
+
+ +

Reflections on 20 Years of Person Re-ID

+

Measuring the problem, watching the field grow, and reconsidering the purpose of the work.

+

VIPeR dataset: download, details, and citation

+
+
+

In 2006, I read Gheissari and colleagues' paper on recognizing people across cameras. The accuracy looked impressive, but the evaluation used just 44 people, seen largely from the front.1 That was a much easier problem than the one I recognized from my undergraduate work.

+ +

I had been working on crowd counting: estimating how many people were in a scene, perhaps 50 or 100 of them moving through it, often seen from an awkward camera angle. People faced different directions, overlapped, and appeared at different scales. A collection of mostly frontal images did not represent that variety.

+ +

The issue was not whether a recognition method could work in the setting that paper studied. It was what the results told us about finding someone from an arbitrary camera angle, with different lighting, a different pose, and hundreds of plausible matches. How could we make progress on the real problem without measuring it properly, even if a realistic test seemed impossibly difficult?

+ +

That question led me to start collecting the Viewpoint Invariant Pedestrian Recognition dataset, VIPeR. The first paper appeared in 2007, but the work began in 2006. Twenty years later, the benchmark has been cited thousands of times and helped researchers develop methods well beyond anything we could do then. It also made surveillance research a promising career path for me. I had an early foothold in a growing field, a dataset others were adopting, and a clear supply of difficult problems to work on. Continuing would have made sense. I eventually chose a different direction.

+ +

The appearance problem between cameras

+ +

Multi-camera surveillance systems were becoming an active research area. Tracking within one camera was already difficult, but expanding coverage introduced another question: when a person appeared in a second camera, was it someone the system had seen before?

+ +

Overlapping fields of view could provide geometric constraints. Nearby cameras could provide timing constraints: someone who left one view might enter another a few seconds later. Those cues were useful, but could become weak or unavailable when cameras were farther apart, their views did not overlap, or the task was to search old footage without knowing when the person would appear.

+ +

Appearance was another source of evidence, and it was the one a human would naturally use. A person looking through video could compare clothing, shape, texture, and what someone was carrying. We needed a model that could make those comparisons despite changes in viewpoint and illumination. Simple color histograms were common, but they lost spatial information and could confuse people in similar clothing. Templates preserved layout but changed dramatically with pose. Existing models offered pieces of a solution; we lacked a reliable way to compare them on the full range of variation.3

+ +

My earlier crowd-counting work had exposed me to people from all angles and camera orientations. A benchmark for recognition needed to include that variation and enough different identities to make matching meaningful. High accuracy on a small, easy evaluation could not tell us which appearance models would work in a camera network.

+ +

Building VIPeR

+ +

My adviser had purchased two cameras and made a point of telling us they were three-CCD cameras, with separate sensors for red, green, and blue. It is easy to take good color video for granted now that we all carry cameras in our phones. Back then, the camera hardware and its low-level details were a noticeable part of the experiment, especially when so much of an appearance model depended on color.

+ +

We ran two cameras at the same time and hired an undergraduate to help capture footage. I built an annotation tool to draw bounding boxes around people and establish correspondences between the views. The tedious part was finding the same person manually in both recordings. That was precisely the task we wanted the technology to help with.

+ +

The final dataset contained two images of each of 632 identities: 1,264 images in total, each normalized to 48 pixels wide by 128 pixels tall. Most pairs differed in viewpoint by at least 90 degrees. The range of identities and the variation between the paired views were what made it useful.2

+ +
+ Twelve VIPeR identity pairs showing the same pedestrians from different camera viewpoints +
Sample identity pairs from VIPeR. Each pair contains one image from each camera. Download the dataset and read its documentation.
+
+ +

The evaluation protocol was another contribution of the PETS paper. Ranking was not new; the 2006 paper that prompted this work also reported ranked matches. Other work treated recognition as a binary same-person/different-person decision and reported ROC curves. We argued for evaluating the actual search task: given a probe image and a gallery of candidates, where does the correct person appear? A low false-positive rate on individual pairs does not directly tell you whether the right match will be near the top of a large gallery.

+ +

We used the cumulative matching characteristic, or CMC: how often does the correct person appear within the first k results? We did not invent CMC, but making ranking and gallery size explicit was an important part of the evaluation we proposed. It connected appearance matching to the task of helping an operator find someone, rather than reducing the problem to classifying isolated pairs. We also showed how to relate the curve to simplified tracking and reacquisition scenarios.2

+ +

To compare methods consistently, we prescribed 5x2 cross-validation: five randomized two-fold rounds, with randomized image order and mirroring, and averaging both gallery/probe directions. This let us study the appearance model separately from the geometric and timing assumptions of a complete tracking system.

+ +

Too difficult to publish?

+ +

Early versions were rejected from CVPR and ICCV before the work found a home at the PETS workshop in 2007. My recollection is that reviewers cited the low accuracy as one of the main flaws: they did not believe the method worked. From my perspective, that difficulty was evidence that we had finally included important parts of the problem. We needed to measure the gap before we could close it.

+ +

The best rank-one baseline in the 2007 PETS paper, a color correlogram with manually localized body regions, reached roughly 5 percent, as read from its CMC plot.2 Our Ensemble of Localized Features, published at ECCV in 2008, reached about 12 percent. ELF used AdaBoost to select useful local color and texture features and combine their comparisons into a learned similarity function.3 It was progress, but clearly left most of the problem unsolved.

+ +

Ranks beyond the first mattered because we envisioned a system assisting a human operator. The 2008 paper reported that ELF placed the correct match within the first 31 results of a 316-image gallery about 70 percent of the time. The point was not simply to report a larger accuracy number. Sorting the gallery could reduce the work of finding the right person while leaving the final decision to an operator.

+ +

We quantified that idea with an expected-search-time model. Assuming an operator could review one image per second without making mistakes, the expected search took 158 seconds with an unsorted gallery and 28.9 seconds with ELF: an 81.7 percent reduction. These were modeled search times, not timings from a user study, but they made the intended benefit explicit.3

+ +

What changed as the field grew

+ +

VIPeR gave researchers a shared test of appearance matching across viewpoints. A few influential methods show how the field progressed from designing better features to learning how to compare them.

+ +

In 2010, Farenzena and colleagues' Symmetry-Driven Accumulation of Local Features, or SDALF, organized color and texture around body regions and their symmetry. Weighted color histograms, stable color regions, and distinctive repeated patches retained more of a person's visual structure while allowing changes in pose and viewpoint. The paper reported roughly 20 percent rank-one accuracy on VIPeR, compared with about 12 percent for ELF.4

+ +

Köstinger and colleagues' 2012 Keep It Simple and Straightforward MEtric, or KISSME, made metric learning computationally simple. Given pairs labeled as the same or different people, it modeled their feature differences and derived a distance metric from a likelihood-ratio test. Some variation is normal between two views of one person; other variation distinguishes identities. KISSME supplied an efficient, reusable way to learn that distinction. Its VIPeR result was 19.6 percent at rank one, averaged over 100 runs.5

+ +

The next few years brought more careful treatment of distinctive details and misalignment. Zhao, Ouyang, and Wang's 2013 salience methods emphasized patches that distinguished one person from others, rather than treating every region equally. Their unsupervised eSDC feature combination reached 26.74 percent rank one; supervised SalMatch reached 30.16 percent.1617 In 2014, Yang and colleagues' Salient Color Names work combined color-name distributions and color histograms across color spaces, reaching 37.8 percent with its fused representation.18

+ +

Liao and colleagues' 2015 work paired Local Maximal Occurrence features with Cross-view Quadratic Discriminant Analysis. LOMO pooled local color and texture occurrences across horizontal positions, allowing an informative detail to move as the viewpoint changed. Illumination handling made the representation more stable; XQDA learned a discriminative subspace and comparison metric together. The combination reached 40 percent rank one on VIPeR and became a strong baseline for later work.6

+ +

Progress also came from combining comparison models and using the gallery itself. Paisitkriangkrai and colleagues' 2015 metric ensembles optimized ranking across several features and metrics, reaching 45.89 percent. García and colleagues' DCIA used the context of visually similar gallery images to refine an initial ranking, reaching 63.92 percent with KCCA.1920 That distinction matters: re-ranking a gallery has access to information beyond comparing one probe with one candidate.

+ +

Matsukawa and colleagues' 2016 Gaussian of Gaussian descriptor modeled local appearance distributions using both means and covariances, then described how those distributions varied across a region. Combined with XQDA and multiple color spaces, it reached about 50 percent rank one on VIPeR.7 Rich hand-crafted representations still competed well on a dataset with only 316 training identities per fold.

+ +

That small training set was also the motivation for Zhang, Xiang, and Gong's 2016 discriminative null-space method. It sought a projection that collapsed same-identity training examples together while preserving separation between identities. Its fused-feature result reached 51.17 percent rank one.21

+ +

Meanwhile, DeepReID and subsequent neural methods learned representations and matching jointly.8 Larger datasets made that approach increasingly effective. VIPeR continued to appear in tables, often as a transfer target, but no longer provided enough training data to represent the field's main ambitions. The 2016 survey Person Re-Identification: Past, Present and Future offers a useful account of this transition.9

+ +

Later work asked a different question: could a model trained on other datasets recognize people in VIPeR without training on VIPeR itself? DIMN and DualNorm were already testing that idea in 2019, reporting 51.23 and 59.4 percent respectively.2829 Meanwhile, an expanded OSNet study reached 68 percent after fine-tuning on VIPeR, illustrating why the training setting matters.30

+ +

SNR's style normalization and restitution reported 52.9 percent in 2020; RaMoE's relevance-aware mixture of experts reached 56.6 percent in 2021 under the same five-source setting. ACL reached 66.4 percent in 2022 using a different, four-source pool, and ReNorm reached 69.1 percent in 2024.22232425 In 2025, CLIP-FGDI adapted a pretrained vision-language model using learned prompts and reported 76.0 percent in its main VIPeR comparison.26 These results belong to the benchmark's history, but they are not a continuation of one unchanged experiment.

+ +
+ Scatter plot of reported VIPeR rank-one accuracy from 2007 to 2026, with selected methods labeled and NK3ML's 99.8 percent marked as a reported claim. +
Reported VIPeR rank-one accuracy over time. Training data and evaluation settings vary across methods. Data and paper references.
+
+ +
+ Values for labeled methods + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
Year and methodRank one
Historical benchmark methods
2007 PETS baseline~5%
2008 ELF~12%
2010 SDALF~20%
2012 KISSME19.6%
2013 eSDC26.74%
2013 SalMatch30.16%
2014 SCNCD37.8%
2015 LOMO + XQDA40.0%
2015 Metric ensembles45.89%
2015 KCCA + DCIA (re-ranking)63.92%
2016 GOG + XQDA49.7%
2016 NFST fusion51.17%
2019 OSNet + VIPeR fine-tuning68.0%
Unseen-target domain generalization
2019 DIMN51.23%
2019 DualNorm (ResNet50)59.4%
2020 SNR52.9%
2021 RaMoE56.6%
2022 ACL66.4%
2024 ReNorm69.1%
2025 CLIP-FGDI76.0%
Other reported results
2018 NK3ML (reported claim)99.8%
2026 RND73.8%
+
+ +
+ Google Scholar citation histories: the PETS paper peaks in 2017 and the ECCV paper in 2018 +
Annual citations to the two foundational papers, from their Google Scholar histories observed on October 12, 2026. The papers are shown separately; 2026 is incomplete. Chart data and sources.12
+
+ +

The citation histories reflect that change. The PETS paper's annual citations peaked at 179 in 2017; the ECCV paper's peaked at 213 in 2018. Google Scholar now lists 1,405 and 2,016 citations respectively. New papers still cite them, but the pace has slowed as the field has moved to larger and more varied benchmarks.

+ +

Outgrowing the benchmark

+ +

Market-1501 made the limitations of the older benchmarks explicit: small identity sets, few cameras, hand-drawn boxes, and too little variation to represent deployment. It introduced more than 32,000 detector-produced boxes for 1,501 identities across as many as six cameras.10 MARS extended the task to video tracklets.11 Later datasets added occlusion, infrared imagery, larger galleries, domain shift, and people who were absent from the gallery altogether.

+ +

The dataset has not disappeared. A CVPR 2026 study of lifelong re-identification still includes VIPeR among seven unseen test datasets, alongside five training datasets.27 Its headline unseen-domain results are averages across those datasets, not individual VIPeR scores. That is another way the question has changed: from learning a matcher for one camera pair to retaining useful representations as new domains arrive.

+ +

VIPeR was superseded, which is a good outcome for an early benchmark. It helped establish a measurable problem and then revealed what the next evaluations needed to include. Its small, carefully cropped images remain useful for studying a particular matching task, but high performance there cannot establish how a system behaves in a city, with detection errors, changing clothes, and an unknown population.

+ +

A promising career, and a question I could not set aside

+ +

I started at UC Santa Cruz about a week after September 11, 2001. By the time I began learning computer vision in 2004, security was widely treated as a reason to accept less privacy. Surveillance research attracted funding and energy, and tracking people across cameras seemed like an important technical problem to help solve.

+ +

By the time VIPeR was gaining attention, I had several reasons to stay in the area. I understood the problem, had built a benchmark the community wanted, and could see many avenues for further research. The technology had growing commercial demand as well as academic interest. Doubling down would have been a straightforward way to build on the work I had already done.

+ +

But a few years into the Ph.D., I began questioning the application. Before advancing to candidacy in 2008, I changed my proposed dissertation title to Privacy Preserving Pedestrian Recognition. I was trying to find a way to continue the research while addressing the concern that it would make people easier to track.

+ +

The concern became more concrete through my adviser, Hai Tao, who was building Vion, a company selling surveillance software. My understanding at the time was that its biggest customer was the Chinese government. I could see a path from our work on recognizing people across cameras to systems for tracking people across cities. I did not want that to be the purpose of my research.

+ +

Vion's later public materials give some context for that concern. The company describes applications in electronic policing, public safety, transportation, and smart-city management. A 2024 article credits Hai and me with VIPeR and describes re-identification as reconstructing movement trajectories. It also presents avoiding face, iris, and gait recognition as a privacy benefit.1314 Those materials do not establish where this particular dataset was used. They do illustrate the distinction I was struggling with: avoiding a particular biometric does not answer whether someone should be tracked.

+ +

I could propose technical safeguards, but the decision to preserve privacy depended on policy. Who could search the camera network? Whose movements could be reconstructed? For what purpose, with what oversight, and with what possibility of refusal? Better features or a formal privacy guarantee could answer only part of that. This became one of several reasons I never finished the Ph.D., and probably one of the larger ones.

+ +

The current debate about Flock cameras makes that question feel familiar. Reading a license plate is, at its core, optical character recognition: recovering an explicit identifier that is already printed on a vehicle. It has its own difficult imaging conditions, and Flock also searches vehicle characteristics beyond the plate.15 But I do not walk around the city with a license plate on. Person re-identification tries to make someone searchable across cameras using appearance alone. In that sense it could provide a license-plate-like tracking capability for people who carry no such identifier, with the potential to be even more intrusive.

+ +

Last month, in September 2026, my own neighborhood, Ladera in Portola Valley, voted to end its Flock camera contract because of privacy concerns. That brought the question out of research papers and into the place where I live. The decision was about whether we wanted this form of surveillance in our community, not whether the cameras could read plates more accurately.

+ +

Looking back, I am proud that VIPeR helped people measure a difficult problem instead of avoiding it. That does not resolve my doubts about what we made easier to do. The question that got me started was whether we could recognize people across cameras. The question I could not set aside was what would happen when we could. The vote in Ladera reminded me that this is not something researchers get to decide for the people who live with the cameras.

+ +

VIPeR dataset download, documentation, and citation guidelines

+ +
+

Sources and notes

+
    +
  1. N. Gheissari et al., Person Reidentification Using Spatiotemporal Appearance, CVPR 2006.
  2. +
  3. D. Gray, S. Brennan, and H. Tao, Evaluating Appearance Models for Recognition, Reacquisition, and Tracking, PETS 2007.
  4. +
  5. D. Gray and H. Tao, Viewpoint Invariant Pedestrian Recognition with an Ensemble of Localized Features, ECCV 2008.
  6. +
  7. M. Farenzena et al., Person Re-Identification by Symmetry-Driven Accumulation of Local Features, CVPR 2010.
  8. +
  9. M. Köstinger et al., Large Scale Metric Learning from Equivalence Constraints, CVPR 2012, Table 2(a); authors' code and documentation.
  10. +
  11. S. Liao et al., Person Re-Identification by Local Maximal Occurrence Representation and Metric Learning, CVPR 2015; authors' results.
  12. +
  13. T. Matsukawa et al., Hierarchical Gaussian Descriptor for Person Re-Identification, CVPR 2016.
  14. +
  15. W. Li et al., DeepReID: Deep Filter Pairing Neural Network for Person Re-identification, CVPR 2014.
  16. +
  17. L. Zheng, Y. Yang, and A. G. Hauptmann, Person Re-Identification: Past, Present and Future, 2016. See also Karanam et al.'s systematic benchmark and Ye et al.'s deep-learning survey.
  18. +
  19. L. Zheng et al., Scalable Person Re-Identification: A Benchmark, ICCV 2015.
  20. +
  21. L. Zheng et al., MARS: A Video Benchmark for Large-Scale Person Re-Identification, ECCV 2016.
  22. +
  23. Google Scholar article histories for PETS 2007 and ECCV 2008, observed October 12, 2026. The plotted range is 2010–2026, the common range displayed in the two histories.
  24. +
  25. Vion Technology, article on its re-identification patents and history, 2024 (Chinese).
  26. +
  27. Vion Technology, company history; Beijing Security and Protection Industry Association, company profile (Chinese).
  28. +
  29. Flock Safety, Vehicle Recognition Software for Investigation-grade LPR, 2026; describes plate reading and searches by vehicle appearance.
  30. +
  31. R. Zhao, W. Ouyang, and X. Wang, Unsupervised Salience Learning for Person Re-identification, CVPR 2013, Table 1. The chart uses eSDC ocsvm with feature fusion.
  32. +
  33. R. Zhao, W. Ouyang, and X. Wang, Person Re-identification by Salience Matching, ICCV 2013, Section 5 and Figure 5(a).
  34. +
  35. Y. Yang et al., Salient Color Names for Person Re-identification, ECCV 2014, Table 1. The chart uses Final(ImgF), including color-histogram fusion.
  36. +
  37. S. Paisitkriangkrai, C. Shen, and A. van den Hengel, Learning to Rank in Person Re-identification with Metric Ensembles, CVPR 2015, Figure 1.
  38. +
  39. J. García et al., Person Re-Identification Ranking Optimisation by Discriminant Context Information Analysis, ICCV 2015, Table 2. Includes gallery-context re-ranking.
  40. +
  41. L. Zhang, T. Xiang, and S. Gong, Learning a Discriminative Null Space for Person Re-Identification, CVPR 2016, Table 1, Ours (Fusion).
  42. +
  43. X. Jin et al., Style Normalization and Restitution for Generalizable Person Re-Identification, CVPR 2020, supplementary Table 5. The five-source setting includes DukeMTMC-reID; the alternative four-source setting reports 55.1 percent.
  44. +
  45. Y. Dai et al., Generalizable Person Re-Identification with Relevance-Aware Mixture of Experts, CVPR 2021, Table 2, Protocol 1. Protocol 2 uses different source datasets and reports 63.4 percent.
  46. +
  47. P. Zhang et al., Adaptive Cross-Domain Learning for Generalizable Person Re-Identification, ECCV 2022, Table 1. Uses four source datasets, omitting DukeMTMC.
  48. +
  49. R. Nie et al., Rethinking Normalization Layers for Domain Generalizable Person Re-Identification, ECCV 2024, Table 1.
  50. +
  51. H. Zhao, L. Qi, and X. Geng, CILP-FGDI: Exploiting Vision-Language Model for Generalizable Person Re-Identification, IEEE TIFS 2025, Table I; authors' implementation. The title spells the acronym CILP-FGDI; the method and code use CLIP-FGDI. The chart uses the main-table 76.0 percent, not the 77.5 percent hyperparameter-ablation result in Table VIII.
  52. +
  53. K. Xu et al., Vision-Language Attribute Disentanglement and Reinforcement for Lifelong Person Re-Identification, CVPR 2026; supplementary Tables 3-4. UnSeen-Avg combines seven datasets and is not a VIPeR rank-one result.
  54. +
  55. J. Song et al., Generalizable Person Re-Identification by Domain-Invariant Mapping Network, CVPR 2019, Table 2.
  56. +
  57. J. Jia, Q. Ruan, and T. M. Hospedales, Frustratingly Easy Person Re-Identification: Generalizing Person Re-ID in Practice, BMVC 2019, Table 5. The chart uses ResNet50 + DualNorm; the main MobileNetV2 configuration reports 53.9 percent.
  58. +
  59. K. Zhou et al., Learning Generalisable Omni-Scale Representations for Person Re-Identification, October 2019 preprint, Table 4. Includes VIPeR fine-tuning after external re-identification pretraining.
  60. +
+
+
+
+ +
+ + + + diff --git a/site/writing/viper-nearly-twenty-years-later/index.html b/site/writing/viper-nearly-twenty-years-later/index.html new file mode 100644 index 0000000..2de0288 --- /dev/null +++ b/site/writing/viper-nearly-twenty-years-later/index.html @@ -0,0 +1,13 @@ + + + + + + + + Reflections on 20 Years of Person Re-ID + + +

Reflections on 20 Years of Person Re-ID

+ +