Evaluation of Deep Learning Instance Segmentation Models for Pig Precision Livestock Farming

In this paper, the deep learning instance segmentation architectures DetectoRS, SOLOv2, DETR and Mask R-CNN were applied to data from the field of Pig Precision Livestock Farming to investigate whether these models can address the specific challenges of this domain. For this purpose, we created a custom dataset consisting of 731 images with high heterogeneity and high-quality segmentation masks. For evaluation, the standard metric for benchmarking instance segmentation models in computer vision, the mean average precision, was used. The results show that all tested models can be applied to the considered domain in terms of prediction accuracy. With a mAP of 0.848, DetectoRS achieves the best results on the test set, but is also the largest model with the greatest hardware requirements. It turns out that increasing model complexity and size does not have a large impact on prediction accuracy for instance segmentation of pigs. DETR, SOLOv2, and Mask R-CNN achieve similar results to DetectoRS with a parameter count almost three times smaller. Visual evaluation of predictions shows quality differences in terms of accuracy of segmentation masks. DetectoRS generates the best masks overall, while DETR has advantages in correctly segmenting the tail region. However, it can be observed that each of the tested models has problems in assigning segmentation masks correctly once a pig is overlapped. The results demonstrate the potential of deep learning instance segmentation models in Pig Precision Livestock Farming and lay the foundation for future research in this area.