Automated Landmark-based Cat Facial Analysis and its Applications

This paper systematically explores the utility of an automated 48-landmark detector for cat facial analysis, addressing the labor-intensive nature of manual annotation. The study developed AI pipelines for three benchmark tasks using two previously collected datasets: cat breed recognition, cephalic type recognition, and pain recognition. While replacing manual landmarks with automated ones generally decreased performance across all tasks, this reduction was an acceptable trade-off for full automation in some areas. Specifically, automated pipelines achieved 75% accuracy in cephalic type recognition and 66% in pain recognition, suggesting landmark-based approaches show promise for automated pain assessment and morphological explorations. However, the breed recognition pipeline performed poorly, indicating that deep learning approaches using direct image data might be more suitable for this task. The study emphasizes the need for valid benchmarks and publicly accessible datasets in animal facial analysis, similar to human affective computing, while also considering ethical implications.