Computer vision researcher · BSc Computer Science & Engineering, TU Delft
Keypoints, matching and camera pose, mostly where GPS does not reach: drones, the Moon, other people's models.
|
🏔️ ETH Zürich Computer Vision and Geometry Group research intern · since July 2026 |
Prof. Marc Pollefeys' group, supervised by Philipp Lindenberger. Keypoint detection and uncertainty for more reliable camera pose: a new training objective for a learned detector, built on the group's RaCo (3DV 2026) and evaluated on HPatches, MegaDepth, ScanNet and ETH3D on the Euler cluster. Preprint in preparation. |
|
🌙 ESA ESTEC GNC section April to June 2026 |
Cross-modal feature matching, event camera to optical imagery, for terrain-relative navigation in lunar landings. Test scenarios built in the PANGU simulator. |
|
🛸 Scaled Autonomy computer vision engineer December 2025 to June 2026 |
Drone-to-satellite image registration for GPS-denied flight. LoFTR plus Fourier-Mellin matching, four times the baseline inlier ratio on real flight data. |
|
🎓 TU Delft BSc CSE, multimedia variant GPA 8.0 |
Computer vision and ML track. Real-time multi-camera 3D drone tracking, part-time ML engineer at Dream Team Epoch, former board member of the debating club. |
I left school early to teach myself maths and CS, and got into TU Delft through an independent entrance exam. Tea is always on.
🔊 sound
no-swim-back-42.mp4
🐦⬛ ARC-AGIThe Abstraction and Reasoning Corpus: tiny grid puzzles that humans solve at a glance and machines still find maddening. I trained a small GPT-2 from scratch on grids serialised into a 14-token vocabulary, with D8 symmetry and colour-permutation augmentation. It learns the textures and gets the grid shape wrong, which is the whole problem in one line. |
TU Delft image processing project, team lead. Python/OpenCV pipeline: plate localisation, character segmentation, SVM recognition, automated evaluation in CI. |
|
Visual model surgery for CNNs. Load a PyTorch model, see its weights as visual mass, click a layer, ablate it, fork the model, and diff what changed: in the prediction and on the graph. |
Camera calibration from scratch: Direct Linear Transform with Hartley normalisation and non-linear refinement. Just Python and NumPy, no black boxes. |
🕰️ Currently brewing (private, for now)
- keypoint-gap-agent: autonomous gap hunter for keypoint detectors on edge hardware. Cloud agents scan the literature every evening; Euler jobs run the experiments overnight.
- event-trn: event-camera terrain-relative navigation, the ESA work, being written up.
- geo-loftr-KD: knowledge distillation for geometry-aware LoFTR matching.
🎩 “We're all mad here.”