WildMatch: Weakly Supervised Image Matcher Adaptation for Wildlife Re-Identification
WildMatch adapts a pretrained keypoint matcher for animal re-ID using only identity labels, not keypoint annotations.
The method mines useful image pairs, treats matching identities as weak positives and different identities as weak negatives, then fine-tunes the matcher contrastively. The authors report gains over off-the-shelf matchers and a local-global fusion baseline across open wildlife re-identification datasets. In held-out-individual tests, they say the model learns transferable correspondences rather than memorizing known animals. HF Daily Papers' note
The method mines useful image pairs, treats matching identities as weak positives and different identities as weak negatives, then fine-tunes the matcher contrastively. The authors report gains over off-the-shelf matchers and a local-global fusion baseline across open wildlife re-identification datasets. In held-out-individual tests, they say the model learns transferable correspondences rather than memorizing known animals. HF Daily Papers' note
score 4