Published September 19, 2025
A masked ViT for person re-identification
Computer VisionViTRe-Identification
Re-ID in one line
Given a photo of a person, the task is to find the same person in a gallery of other images.
The mask trick
The background adds identity features that generalize poorly. For that reason the ViT receives a segmentation mask together with the image, which focuses the model's attention on the silhouette. This has two effects:
- fewer false matches caused by similar backgrounds;
- better robustness to changes in viewpoint and lighting.
The demo on this site runs on synthetic data, and no real biometric images are processed.