Before and after fine-tuning¶
What to look for: the same two photos, scored twice. Fine-tuning raises the image score on every pair, from below 0.42 to above 0.60, and multiplies the correspondences on the animal.
The most direct view of what WildMatch changes: one photo pair, the query and its correct top-1 gallery photo of the same individual, matched by the default LoMa matcher and by the matcher fine-tuned on that dataset. Switch between the two and watch the correspondences and the image score change. Matches are computed on the background-removed inputs the paper uses and drawn on the original photos; the pairs are the ones in the paper's qualitative figure.
- Matcher toggle: the default checkpoint against the checkpoint fine-tuned on the shown dataset; each button carries that matcher's score and match count.
- Pair selector: one query and top-1 pair for each of the eight datasets in the paper.
- Slider: how many of the strongest correspondences are drawn; hover a line for its confidence.
This is the only demo that shows the default matcher next to the fine-tuned one. Scores here can differ from the paper's recorded values by up to 0.02, because the demo extracts features one image at a time while the benchmark extracts them in batches.
Photographs from CzechLynx (Picek et al.), WildlifeReID-10k and SalamanderID2025 (AnimalCLEF 2025), the latter shown with the dataset team's permission.