Skip to content

Demo

Five interactive views of what WildMatch does, each on its own page. They follow the paper's argument: first the direct before/after comparison, then an honest look at where fine-tuning helps and where it does not, then the weak supervision the method learns from, the background-masking step every image passes through, and finally a playground on synthetic renders. Use the arrows at the bottom of each page to walk through them in order.

Try it on your own photos

The WildMatch Space on Hugging Face matches two photos of your choice with the default or the fine-tuned matcher, and finds a lynx in the gallery of 44 individuals the matcher never saw.

How to read the demos

  • 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.
  • Lines are correspondences between the two photos. Their weight follows confidence, hovering one shows its exact value, and a slider sets how many of the strongest are drawn.
  • Masks: matches are computed on background-removed inputs, the inputs the paper uses, and drawn on the original photos.
  • Frames: blue marks the true individual, red marks a hard negative or a SAM 3 mask outline; each page explains its own colours.

Photographs from CzechLynx (Picek et al.), WildlifeReID-10k and SalamanderID2025 (AnimalCLEF 2025), the latter shown with the dataset team's permission. Synthetic lynx renders from the CzechLynx synthetic subset, Zenodo record 17592004, CC BY 4.0.