A segment-based framework for explainability in animal affective computing

Recent developments in animal motion tracking and pose recognition have revolutionized the study of animal behavior. More recent efforts extend beyond tracking towards affect recognition using facial and body language analysis, with far-reaching applications in animal welfare and health. Deep learning models are the most commonly used in this context. However, their "black box" nature poses a significant challenge to explainability, which is vital for building trust and encouraging adoption among researchers. Despite its importance, the field of explainability and its quantification remains under-explored. Saliency maps are among the most widely used methods for explainability, where each pixel is assigned a significance level indicating its relevance to the neural network's decision. Although these maps are frequently used in research, they are predominantly applied qualitatively, with limited methods for quantitatively analyzing them or identifying the most suitable method for a specific task. In this paper, we propose a framework aimed at enhancing explainability in the field of animal affective computing. Assuming the availability of a classifier for a specific affective state and the ability to generate saliency maps, our approach focuses on evaluating and comparing visual explanations by emphasizing the importance of meaningful semantic parts captured as segments, which are thought to be closely linked to behavioral indicators of affective states. Furthermore, our approach introduces a quantitative scoring mechanism to assess how well the saliency maps generated by a given classifier align with predefined semantic regions. This scoring system allows for systematic, measurable comparisons of different pipelines in terms of their visual explanations within animal affective computing. Such a metric can serve as a quality indicator when developing classifiers for known biologically relevant segments or help researchers assess whether a classifier is using expected meaningful regions when exploring new potential indicators. We evaluated the framework using three datasets focused on cat and horse pain and dog emotions. Across all datasets, the generated explanations consistently revealed that the eye area is the most significant feature for the classifiers. These results highlight the potential of the explainability frameworks such as the suggested one to uncover new insights into how machines 'see' animal affective states.