TRADE-OFF BETWEEN ACCURACY, VISUAL QUALITY, AND COMPUTATIONAL COST IN CATTLE SEGMENTATION: AN EVALUATION OF MASK R-CNN, POINTREND, AND U-NET

Accurate Body Condition Score (BCS) assessment is essential for precision livestock farming, but traditional methods are subjective. This work compared three deep learning architectures (Mask R-CNN, Mask R-CNN with PointRend and U-Net) to segment cattle outlines in a slaughterhouse, using 130 manually annotated images. The evaluation focused on pixel-level metrics and computational cost. The results showed that PointRend achieved the best overall pixel-by-pixel performance, as well as having the best visual quality of the masks and the greatest adherence to the real silhouette. U-Net, on the other hand, although it has achieved identical results to Mask R-CNN, can present severe limitations, such as merging adjacent instances, fragmentation of masks, and significantly higher training time. The choice of model must consider the trade-off between quantitative accuracy, contour quality, robustness in noisy environments, and computational cost for refrigerator applications.