A masked ViT for person re-identification

Re-ID in one line

Given a query image of a person, find the same person across a gallery of shots.

The masking trick

Backgrounds leak identity cues that don't generalize. We feed the ViT a segmentation mask so attention concentrates on the silhouette:

• fewer false matches from similar backgrounds
• better robustness to viewpoint and lighting

Demo on the site runs on synthetic data — no real biometric images are processed.