Going Deeper than Tracking: A survey of computer-vision based recognition of animal pain and emotions

This study offers a comprehensive survey of computer vision-based research aimed at recognizing internal affective states in animals, such as pain and emotions, to enhance animal welfare. It specifically focuses on facial and bodily behavior analysis in mammals, extending beyond mere motion tracking. The survey highlights significant challenges, including the lack of a common definition for animal emotions, the inherent difficulty in establishing ground truth due to the absence of verbal communication, data scarcity, and morphological diversity across species. Methodological aspects covered include data collection practices (e.g., recording equipment and environment), annotation strategies (distinguishing between stimulus-based and behavior-based annotations), and data analysis approaches such as using single frames versus spatio-temporal sequences, parts-based versus holistic methods, and hand-crafted versus learned features (deep learning). The authors discuss the “black-box” nature of deep learning models, which can impede explainability crucial for clinical and welfare applications. To advance the field, the paper proposes best practice recommendations for handling data imbalance, applying robust cross-validation techniques (like leave-one-animal-out to ensure subject exclusivity), and exploring domain transfer methodologies. A critical future direction emphasized is the need for public, benchmark datasets to enable standardized comparisons and foster progress in this interdisciplinary area.