GorillaWatch: An Automated System for In-the-Wild Gorilla Re-Identification and Population Monitoring

TL;DR

Automating gorilla re-identification and population monitoring from camera-trap footage is hampered by manual labeling and a lack of in-the-wild datasets. The authors present GorillaWatch, an end-to-end pipeline augmented with multi-frame self-supervised pretraining and differentiable explainability, evaluated on three new benchmarks: Gorilla-SPAC-Wild, Gorilla-Berlin-Zoo, and Gorilla-SPAC-MoT. Key contributions include a comprehensive open-world dataset suite, temporal pretraining that improves re-ID by up to 11%, a differentiable AttnLRP-based faithfulness check, PTAM for tracklet aggregation, and constrained clustering to improve population counting. Results show that aggregating features from large-scale image backbones with temporal pretraining outperforms specialized video architectures in data-scarce wildlife settings, enabling scalable, non-invasive monitoring and providing a practical foundation for conservation analytics; all code and datasets are released for community use.

Abstract

Monitoring critically endangered western lowland gorillas is currently hampered by the immense manual effort required to re-identify individuals from vast archives of camera trap footage. The primary obstacle to automating this process has been the lack of large-scale, "in-the-wild" video datasets suitable for training robust deep learning models. To address this gap, we introduce a comprehensive benchmark with three novel datasets: Gorilla-SPAC-Wild, the largest video dataset for wild primate re-identification to date; Gorilla-Berlin-Zoo, for assessing cross-domain re-identification generalization; and Gorilla-SPAC-MoT, for evaluating multi-object tracking in camera trap footage. Building on these datasets, we present GorillaWatch, an end-to-end pipeline integrating detection, tracking, and re-identification. To exploit temporal information, we introduce a multi-frame self-supervised pretraining strategy that leverages consistency in tracklets to learn domain-specific features without manual labels. To ensure scientific validity, a differentiable adaptation of AttnLRP verifies that our model relies on discriminative biometric traits rather than background correlations. Extensive benchmarking subsequently demonstrates that aggregating features from large-scale image backbones outperforms specialized video architectures. Finally, we address unsupervised population counting by integrating spatiotemporal constraints into standard clustering to mitigate over-segmentation. We publicly release all code and datasets to facilitate scalable, non-invasive monitoring of endangered species

Figures (7)

• Figure 1: Overview of our complete re-identification pipeline.
• Figure 2: Sample images from the Gorilla-SPAC-Wild dataset (top row) and the Gorilla-Berlin-Zoo dataset (bottom row).
• Figure 3: Zero-shot performance comparison of different pre-pretrained models on the Gorilla-SPAC-Wild test set.
• Figure 4: Top-1 accuracy of DINOv2Giant after supervised fine-tuning on the Gorilla-SPAC-Wild training set.
• Figure 5: Effect of Multi-frame Pretraining: We evaluate performance across single-frame, 4-frame, and 10-frame temporal sequences. The green bars denote accuracy immediately after pretraining, while the brown bars indicate the accuracy change ($\Delta$) after fine-tuning.