Anchored Graph Clustering with Local Feature Matching for AnimalCLEF 2026

AnimalCLEF 2026 evaluates individual animal discovery and re-identification across four species: Eurasian lynx, fire salamander, loggerhead sea turtle, and Texas horned lizard. Submissions assign test images to identity clusters and are evaluated using the Adjusted Rand Index (ARI) against hidden ground-truth identities. We present a per-species anchored graph-clustering framework combining pretrained re-identification descriptors (MiewID and MegaDescriptor), LightGlue local feature matching, and a tabular pairwise classifier for refinement edges. The system anchors the clustering graph to the labelled reference set: when a test image's blended similarity to a known training individual exceeds a species-specific threshold, it inherits that identity anchor. Test images sharing the same anchor are grouped together, while non-anchored pairs are merged only when supported by a high-confidence pairwise classifier score. Our selected submission achieved 0.61741 ARI on the public leaderboard and 0.57038 ARI on the private leaderboard.