Faces, Flukes, Fins, and Flanks: How Multispecies Re-ID Models are Transforming Our Approach to AI..
summary
Wildbook deploys production pipelines that already support ~1,900 field biologists tracking just over 221,000 individuals from >1,000,000 encounters and 10,000,000 photos; teams get a ranked list with cosine scores per query and don’t need model retraining to add new animals. Moving from species-specific matchers to multispecies embeddings (ArcFace-style training, ~2,052-dim vectors) yields faster days-to-train cycles, a 54-species/61-class model that generalizes, and clustering that recovers ~70% of samples at ≥95% precision for cold-start catalogs.
AI-written summary · about
chapters (8)
- 0:00Wild Me and Conservation AI
- 2:31Conservation needs AI engineering
- 8:54Computer vision in wildlife monitoring
- 17:57Single-species to multispecies AI
- 22:58Open-world re-ID robustness
- 30:51Generalized AI agents for conservation
- 41:59Q&A on training and data
- 57:41Closing remarks and future
related talks
44mWildbook: Software and Machine Learning Engineers Join the Fight...
39mOverview and Marine Bioacoustics and Perch Model
39mBuilding Highly Scalable Facial Recognition Pipelines
48mSetting the stage for AI for biodiversity
30mIn the service of the history. AI in archivistics.
01h 23mMultimodal Deep Learning in the Real World
30mThe Art and Science of Deep Face Recognition in Python
33mDo I Know You? Real time facial recognition with an Apache stack