Towards Application-Specific Evaluation of Vision Models: Case Studies in Ecology and Biology

TL;DR

This work foregrounds the need for application-specific evaluation in vision models for ecology and biology, arguing that ML metrics alone can misrepresent downstream usefulness. It presents two case studies: (i) chimpanzee abundance/density estimation via camera trap distance sampling (CTDS) when filtering video with a behaviour classifier, and (ii) head-rotation-based gaze estimation in pigeons using 3D posture estimators. The findings show that a high ML performance (e.g., $mAP$ near $87.8\%$) does not guarantee accurate downstream inferences, as automated pruning can still bias abundance estimates and the best Euclidean/keypoint accuracy does not necessarily yield the most reliable gaze estimates. The authors advocate integrating application-specific metrics into datasets and benchmarking pipelines, enabling evaluation that aligns with real-world ecological/biological workflows, and call for closer interdisciplinary collaboration to bridge CV advances with end-user needs.

Abstract

Computer vision methods have demonstrated considerable potential to streamline ecological and biological workflows, with a growing number of datasets and models becoming available to the research community. However, these resources focus predominantly on evaluation using machine learning metrics, with relatively little emphasis on how their application impacts downstream analysis. We argue that models should be evaluated using application-specific metrics that directly represent model performance in the context of its final use case. To support this argument, we present two disparate case studies: (1) estimating chimpanzee abundance and density with camera trap distance sampling when using a video-based behaviour classifier and (2) estimating head rotation in pigeons using a 3D posture estimator. We show that even models with strong machine learning performance (e.g., 87% mAP) can yield data that leads to discrepancies in abundance estimates compared to expert-derived data. Similarly, the highest-performing models for posture estimation do not produce the most accurate inferences of gaze direction in pigeons. Motivated by these findings, we call for researchers to integrate application-specific metrics in ecological/biological datasets, allowing for models to be benchmarked in the context of their downstream application and to facilitate better integration of models into application workflows.