AI camera-trap retrieval benchmark achieves 34% F1, doubling prior models

Automatically retrieving specific events from massive camera-trap video datasets remains a challenge for ecologists. A team of researchers (EPFL, Université de Lausanne, and others) has developed Prompting-MammAlps—the first dedicated benchmark for text-to-video retrieval (TVR) in ecological camera-trap footage. The benchmark includes 775 candidate videos and 135 ecologically relevant queries, designed to test fine-grained spatiotemporal understanding. Existing large video-language models (VLMs) often fail on such data due to poor generalization and lack of interpretability.

To address this, the team built a two-stage retrieval pipeline. First, a vision transformer (ViT) performs spatiotemporal action localization on each video, producing a structured textual description of detected events (e.g., “a deer walking left, then stopping”). Second, an LLM-based coding agent, guided by ethology-inspired queries, parses this structured text using a custom parsing library—minimizing hallucination risks and improving interpretability. The approach achieved a set-based F1-score of 34%, nearly double the 18% F1 of the best zero-shot VLM. The work was accepted at ECCV 2026 and is open-source.