AnimalCLEF 2026

Solo entry to the AnimalCLEF 2026 individual animal re-identification challenge (LifeCLEF Lab, CLEF 2026). Final rank: 6/230 on the private leaderboard, top solo team.

The challenge

AnimalCLEF 2026, the individual animal re-identification task at the LifeCLEF Lab of CLEF 2026, ran from 1 February to 7 May 2026 and drew 230 teams. Competitors must assign every test image to an identity cluster across four species: Eurasian lynx, fire salamander, loggerhead sea turtle, and Texas horned lizard. The test set mixes known training individuals with completely unseen ones; submissions are scored by Adjusted Rand Index (ARI), which measures pairwise cluster consistency against hidden ground-truth identities. Both false merges and missed merges hurt, and novel individuals must be assigned to a cluster; the scoring gives no credit for leaving them isolated.

Each species demands something different. Lynx coat patterns and turtle scute arrangements separate cleanly on global appearance descriptors; salamander spot patterns and lizard ventral markings need local feature matching to tell individuals apart at fine scale. The test set spans 2,409 images (946 lynx, 689 salamander, 500 sea turtle, 274 horned lizard); the mix of known and novel individuals varies by species, which shapes the pipeline design for each.

Result

6th of 230 teams on the private leaderboard, first place among solo entries. Final scores: 0.61741 ARI public / 0.57038 ARI private, producing 833 predicted clusters across 2,409 test images.

System diagram for the AnimalCLEF 2026 entry

The pipeline

Per-species anchored graph-clustering pipeline. Global similarity comes from two pretrained re-identification encoders: MiewID (used off-the-shelf) and MegaDescriptor-L-384, with two species-specific MegaDescriptor checkpoints fine-tuned with ArcFace heads on the competition’s Lynx and Sea Turtle training splits. LightGlue + SuperPoint local feature matching provides a complementary score. The two signals are blended into a per-species similarity, thresholded, and turned into edges in an undirected graph; connected components become the predicted clusters.

The system anchors the graph through the labelled training set: when a test image’s nearest training neighbour exceeds the species threshold, the test image inherits that individual’s identity, and all test images sharing an anchor merge automatically. For Salamander and Horned Lizard, an XGBoost pair classifier scores every test–test pair from LightGlue match statistics and descriptor cosines; pairs above the species threshold contribute must-link edges. The Salamander classifier draws on embeddings from a LoRA-finetuned DINOv2-S adapter (rank 8, trained on ~7,800 hand-labelled spot-correspondence pairs); the Horned Lizard classifier additionally uses DINOv3 patch features. MegaDescriptor fine-tunes use ArcFace heads.