WhaleID

Built with Dr. Caine Delacy and a network of local partners working across humpback whale hotspots in the Indian, Pacific, and Atlantic basins. The thesis is plain: effective protection starts with the ability to track whales as individuals, not as a population in aggregate.

The pipeline is computer vision built around an embedding model and similarity search. Re-identification reduces to a constant-time lookup against the catalog. Inputs are full-body underwater photographs — head tubercle patterns, dorsal ridges, pectoral scarring, fluke geometry, peduncle texture — not isolated fluke patches. Cross-population matches are ranked by a multiplicative fusion score and tiered into four outcomes: auto-merge, known individual, needs review, or new individual.

The pipeline runs five stages: detection (Roboflow YOLO), segmentation (SAM2 Hiera Large), orientation normalization (OpenCV PCA), multi-model feature extraction (MIEW-ID v3, MegaDescriptor-L, ALIKED + LightGlue, NeuralWhale shape and texture descriptors), and a conservative decisioning layer that prioritizes catalog integrity over false confidence. Every stage produces auditable artifacts — masks, overlays, keypoint visualizations — so an expert reviewer can see exactly how a match was made.

The architecture is meant to run locally. Field biologists, conservation NGOs, and community monitors can upload imagery, receive identifications, and contribute to a shared catalog without cloud infrastructure, proprietary tools, or outside control over the data.

The model is not the contribution. The contribution is conservation infrastructure for longitudinal study at population scale across multiple hotspots simultaneously. A co-authored paper on the methodology is expected in 2026.