GorillaWatch: self-supervised in-the-wild gorilla re-identification pipeline (WACV 2026)
gdm3000 · x · 2026-11-05
An HPI-led team released GorillaWatch, an end-to-end system for re-identifying critically endangered western lowland gorillas and monitoring their populations in the wild—featured in National Geographic Germany's October issue and deployed for gorilla conservation in the Republic of the Congo with SPAC and Conservation X Labs.
Key contributions (WACV 2026):
- Three new datasets: Gorilla-SPAC-Wild (the largest wild primate re-identification video dataset to date), Gorilla-Zoo-Berlin (cross-domain generalization), and Gorilla-SPAC-MoT (multi-object tracking in camera-trap footage)
- An end-to-end pipeline integrating detection, tracking, and re-identification
- Multi-frame self-supervised pretraining that exploits tracklet consistency to learn domain-specific features without manual labels
- A differentiable adaptation of AttnLRP verifying the model relies on discriminative biometric traits rather than background correlations
Code, datasets, and arXiv links are all public.
Related event: HPI's AI Gorilla Monitoring System Featured by National Geographic(2 posts)→
More from Research
- Cohere Labs Open Science community to host Kernel Autoresearch and Context Language Models talks — Cohere_Labs · 2026-11-05
- RoboColosseum, a Major In-the-Wild Robotics Benchmark, Picks Unitree G1 for Humanoid Track — chris_j_paxton · 2026-11-05
- Variable-geometry cellular automata: rules that reshape cells and local geometry — snikolov · 2026-11-05
- DialFork: first benchmark targeting procedural hallucination in LLM agents, from Idiap — sergbur · 2026-11-05
- New weight-space benchmark ranks GPT-6 Astra first at 96.3 for writing neural net weights from scratch — BorisMPower · 2026-11-05
- Max Planck paper: reasoning makes LLMs agree more — even on shared mistakes — rohanpaul_ai · 2026-11-05