CoDU — Complex Document Understanding

Document understanding code for the 2025 Samsung AI Challenge. The pipeline combines DocLayout-YOLO layout detection, PaddleOCR text recognition, Pix2Tex formula recognition, bounding-box post-processing, and reading-order reconstruction.

Files

• script.py: document inference and CSV/visualization output.
• model/order/order_again.py: reading-order post-processing.
• model/order/dacon_test.py: local evaluation utilities.
• model/paddleocr/: OCR model configurations.

Setup

Use Python 3.10+ in a GPU environment. The original requirements target PaddlePaddle GPU with CUDA 12.6; install a compatible PyTorch build for your environment. PDF/PPTX conversion also needs Poppler/LibreOffice, respectively.

python -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt pdf2image scikit-learn

Model weights

Download the files below individually and save each at the specified path. Create missing folders; for PaddleOCR downloads, rename the file to the exact filename shown in Save as. The OCR configuration files are already included in this repository.

Model	Download	Save as (relative to repository root)
DocLayout-YOLO	doclayout_yolo_docstructbench.pt	model/yolo/doclayout_yolo_docstructbench.pt
Inline formula detection	formula_yolo.pt	model/yolo/formula_yolo.pt
PaddleOCR detection	paddleocr_det_inference.pdiparams	model/paddleocr/det/PP-OCRv5_server_det/inference.pdiparams
PaddleOCR recognition	paddleocr_rec_inference.pdiparams	model/paddleocr/rec/korean_PP-OCRv5_mobile_rec_infer/inference.pdiparams
Pix2Tex	weights.pth	model/pix2tex/weights.pth
Pix2Tex image resizer	image_resizer.pth	model/pix2tex/image_resizer.pth

Pix2Tex weights enable formula transcription; the script can continue without Pix2Tex if its initialization fails.

File checksums: SHA256SUMS. Datasets are not included.

Run

Prepare data/test.csv with columns ID,path,width,height. Document paths are resolved relative to the CSV directory. Run from the repository root:

python script.py

Results are saved to output/submission.csv and output/viz_45/. Paths and inference settings can be adjusted near the top of script.py.

This repository contains inference code. The train.py referenced in the separate training guide is not included. Source syntax has been checked; full GPU inference has not been validated for this upload.