DS@GT ARC at AnimalCLEF 2026: Species-Aware Graph Construction for Multi-Species Animal Re-Identification

Technical Summary: DS@GT ARC at AnimalCLEF 2026

Problem Definition
The paper addresses the challenge of automated individual animal re-identification (ARC) within a multi-species clustering setting. Unlike previous open-set tasks that required distinguishing known from unknown individuals, the AnimalCLEF 2026 challenge focused on clustering images of four distinct species—Eurasian lynx (Lynx lynx), fire salamander (Salamandra salamandra), loggerhead sea turtle (Caretta caretta), and Texas horned lizard (Phrynosoma cornutum)—into identity groups without prior labels for the test set. The primary difficulty lies in separating identity cues from nuisance variations such as pose, illumination, background, resolution, and species-specific morphology. A critical failure mode identified is the "bridge edge" problem in clustering: high-scoring false positive pairs that, when accepted as graph edges, merge distinct individuals through transitive closure, severely degrading the Adjusted Rand Index (ARI).

Methodology
The DS@GT team formulated re-identification as a species-aware graph construction problem rather than a simple nearest-neighbor retrieval task. The pipeline operates in five distinct stages:

1. Species-Specific Preprocessing: The system applies tailored preprocessing for each species to mitigate domain shifts. This includes vertical-flip test-time augmentation (TTA), SAM3-based foreground masking for salamanders to remove noisy backgrounds, and square padding for lynx images.
2. Global Candidate Retrieval: The system utilizes global descriptors to generate a shortlist of candidate pairs. It evaluates two primary backbones: MegaDescriptor (trained on animal re-identification) and MiewID (a multi-species embedding model). Retrieval budgets are managed to reduce the quadratic complexity of pair comparisons.
3. Local Verification: Shortlisted pairs undergo local feature matching using LightGlue with multiple keypoint families (SuperPoint, SIFT, ALIKED, and DISK). This step provides evidence regarding specific markings, spots, and textures that global descriptors may miss.
4. Learned Pair Scoring: A LightGBM gradient-boosted decision tree model scores candidate pairs. The model integrates global similarity, retrieval rank, neighborhood context features (e.g., shared neighbors, Jaccard overlap), and local verification summaries. Crucially, these scores are treated as edge-strength proposals rather than final identity predictions.
5. Conservative Graph Construction & Clustering: The system employs strict edge admission gates (thresholds, local support requirements, and neighborhood consistency checks) to filter out potential bridge edges. The retained graph is then clustered using Leiden community detection, which allows for splitting weighted components, or connected components analysis.

Key Contributions

• Species-Aware Graph Construction: The paper demonstrates that a single global configuration is insufficient for diverse species. The winning strategy involved selecting specific operating points (backbones, preprocessing, and graph thresholds) for each species based on ablation studies.
• Mitigation of Bridge Edges: The system explicitly addresses the transitive closure failure mode by prioritizing conservative edge admission. The authors show that preventing false merges is more critical than maximizing recall in the initial retrieval stage.
• Integration of Multi-Modal Evidence: The approach successfully fuses global descriptors, local feature matching (LightGlue), and neighborhood context into a unified scoring framework, showing that local evidence is essential for validating global similarities.
• Exploratory 3D and Scale Analysis: The paper documents exploratory work on 3D pose normalization (UV unwrapping) for lynx and scale detection for Texas horned lizards. While these did not make the final submission, they highlight the potential for species-specific geometric cues.

Results
The DS@GT submission achieved a public ARI of 0.733 and a private ARI of 0.674, ranking 5th out of 230 teams.

• Performance Gaps: The top five teams were separated by a narrow margin (0.030 ARI), while the baseline systems (WildFusion and MegaDescriptor) scored significantly lower (0.21 ARI).
• Species-Specific Findings:
  • Salamanders: The most sensitive to preprocessing and backbone choice. Replacing MiewID with MegaDescriptor and removing SAM3 masking significantly reduced performance.
  • Sea Turtles: Stronger retrieval (MegaDescriptor + MiewID fusion) did not automatically translate to better clustering; downstream graph controls were the limiting factor.
  • Lynx: Showed a divergence between public and private scores based on graph shape; larger connected components improved private scores but hurt public scores, indicating different split characteristics.
  • Texas Horned Lizards: Operated as a pure clustering challenge with no labeled training identities, relying on unsupervised graph construction.

Significance and Claims
The paper claims that robust wildlife re-identification requires more than just strong visual representations; it necessitates the calibrated integration of global similarity, local identity markings, neighborhood context, and graph-level constraints. The authors emphasize that the "operating point" of the graph construction—specifically the trade-off between fragmentation and over-merging—is as critical as the descriptor quality itself. The results suggest that for multi-species clustering, a one-size-fits-all approach fails, and species-specific tuning of the entire pipeline (from preprocessing to graph thresholds) is essential for state-of-the-art performance.