WildMatch: Weakly Supervised Image Matcher Adaptation for Wildlife Re-Identification

WildMatch. Weakly Supervised Image Matcher Adaptation for Wildlife Re-Identification. By Turhan Can Kargin and five others, from Jagiellonian University, the Jagiellonian Center for Artificial Intelligence, and one other institution. Published on the fifth of October. Its code repository is worked out: about three hundred source files, with a readme and a licence.

Telling one individual animal from another in camera-trap images is central to monitoring wildlife without disturbing it. Framed as a task, it is instance retrieval: a query image has to find the right individual among a reference set of animals already known. Doing that well means recognising the distinctive local patterns in fur, in skin, or in other visual markings. Current approaches fall into two camps, and each falls short in its own way. One learns global embeddings as a classification problem. That needs many labelled images for every individual, and it largely ignores local evidence, which is exactly where the identifying patterns live. The other applies off-the-shelf image matchers that know nothing in particular about animals.

Those matchers have been pretrained on large and diverse image collections, so adapting them to wildlife looks like the natural move. The obstacle is the data. Wildlife datasets are small, and they lack annotations at the level of correspondences: nobody has marked which point on one photo matches which point on another.

The study asks whether a pretrained keypoint matcher can be adapted using only identity labels. No keypoint-level ground truth, and no geometric correspondence ground truth either. Just the knowledge of which images show the same animal. Those labels are already present in typical monitoring datasets, which is what makes the question practical.

The procedure has three steps. First, the pretrained matcher itself is used to mine informative pairs of images. Second, weak supervision is derived from identity agreement: a pair showing the same individual counts as a positive, a pair showing different individuals as a negative. Third, the matching network is fine-tuned contrastively, strengthening correspondences for same-identity pairs and suppressing them for pairs of different identities. The matcher never sees where a correspondence should be. It only learns, at the level of whole pairs, whether correspondences ought to exist.

Across open-source wildlife re-identification datasets, the approach improves accuracy over off-the-shelf matchers, and over a state-of-the-art method that fuses local and global information. A second test asks what was actually learned. Under an open-world protocol, where some individuals are held out entirely, the adapted matcher still performs, which indicates that it has learned a transferable prior over correspondences rather than memorising the identities it was trained on.

As far as the authors know, this is the first study of identity-supervised adaptation done at the level of the matcher for animal re-identification. Its appeal is in what it asks of the people collecting the data. Nothing new. Identity labels that monitoring projects already keep turn out to be enough to specialise a general image matcher to wildlife, and to do so efficiently with the small datasets the field actually has.