Visual-prompt model enables one-stage wildlife re-ID with 30.6% mAP
A single-stage model using DINOv2 and MegaDescriptor challenges two-stage pipelines in wildlife re-identification.
Wildlife re-identification typically requires a two-stage pipeline: first detect animals, then match identities. Researchers from the University of Johannesburg and KTH Royal Institute of Technology propose a simpler one-stage end-to-end model that performs detection and identity search simultaneously within a shared latent space.
Their architecture builds on DINOv2 for robust spatial geometry and MegaDescriptor for wildlife-specific embedding. To improve identity matching, they inject prompt re-identification features into latent queries, allowing the detection decoder to focus on the target animal. In preliminary tests, the model achieved a 30.584% mAP compared to 44.89% for state-of-the-art two-stage approaches, while offering end-to-end training and faster inference. The work was accepted at the ECCV 2026 Instance-Level Recognition and Generation Workshop in Malmö, Sweden.
- One-stage model combines detection and re-identification in a single pass using latent space search
- Uses DINOv2 for spatial geometry and MegaDescriptor for wildlife re-identification, enhanced by visual-prompt queries
- Achieves 30.584% mAP vs 44.89% for two-stage systems, with qualitative results showing accurate bounding and identity matching
Why It Matters
This could simplify wildlife monitoring pipelines, reducing computational overhead while maintaining competitive accuracy for conservation applications.