Automated recognition of animal pain often relies on two-dimensional saliency maps, which suffer from severe flickering and instability as subjects move across changing camera viewpoints. To overcome this limitation, this study presents an explainability framework that leverages dense 2D-3D correspondence mapping to project neural network attention from video frames onto a canonical three-dimensional surface prototype. Evaluated on horses across post-surgical, orthopedic, and acute pain conditions, the approach achieved video-level classification F1 scores ranging from 0.67 to 0.80. Crucially, the method provides quantitative metrics that map model focus to specific facial regions, establishing statistically significant correlations between machine attention and expert veterinary pain scores for key features like ears and cheek muscles.
Prototype-based Explainable Deep Learning for Sex Classification of Mountain Gazelles in the Wild
Standard deep learning tools in automated wildlife monitoring often function
