Berlin Computer Vision Group

Lukáš Adam is Associate Professor at the University of West Bohemia in Pilsen, Czech Republic, and one of the organizers of AnimalCLEF, a Kaggle competition co-located with CVPR and CLEF 2026. He will walk us through the competition design, the dataset, and why the problem is harder than it looks.
The task: given an image of an animal, decide whether this individual has been seen before. Models need to handle unseen species, long-tail distributions, and images taken years apart under uncontrolled field conditions. The 2026 edition moves toward unsupervised individual discovery -- no labeled training identities, just clustering -- which removes most of the usual scaffolding.
The competition builds on WildlifeReID-10k, a collection of 36 re-identification datasets covering roughly 140,000 images of over 10,000 individuals across dozens of species. Scoring uses the geometric mean of BAKS and BAUS: accuracy on known individuals and accuracy on rejecting unknowns. A model that calls everything "new" scores 0% on BAKS regardless of BAUS. The metric forces you to solve both halves.
Lukáš published on sea turtle re-identification at WACV 2024, co-released the WildlifeDatasets toolkit, and ran AnimalCLEF 2025, which drew 270 participants. If you are considering entering the competition or working on metric learning for wildlife, this is worth attending.
The competition deadline is in April. Details at kaggle.com/c/animal-clef-2026.