Synthetic keypoint matching¶
What to look for: real matcher output on rendered animals. Exactly one gallery render is the same individual as the query; the matcher usually, not always, ranks it first.
Ten synthetic lynxes, two renders each. Choose one of the ten query renders; LoMa + WildMatch, the matcher fine-tuned on CzechLynx, ranks the ten gallery renders, one per individual, so exactly one is the same animal. Pick a candidate to see its correspondences, drag the slider to show more of them, and hover a match to read its confidence. The matches and scores are real matcher output; the images are synthetic renders, not the paper's test data, and the scores are not a benchmark result. Synthetic coats are generated from a shared texture model, so different individuals can look alike to the matcher and a few queries rank another lynx first by a small margin.
How to read it¶
- Score is the matcher's image score: the summed confidence of the mutual-nearest matches above the threshold, divided by the smaller keypoint count, as in the paper.
- Rank orders the ten gallery renders by that score for the chosen query. The gallery render of the query's own individual carries a red inner frame; hover a thumbnail for its details.
- Lines are the strongest correspondences; their weight follows confidence, and hovering one shows its exact value. Every match is available through the slider.
- Synthetic: the images are rendered lynxes, the paper's test data are camera-trap photographs. See Qualitative for matches on real photos, or the before/after demo for the same matcher on real pairs.
Synthetic lynx renders from the CzechLynx synthetic subset (Picek et al.), Zenodo record 17592004, CC BY 4.0.