Occlusion-Resistant instance segmentation of piglets in farrowing pens using center clustering network

Computer vision enables the development of new approaches to monitor behavior, health, and welfare animals. Instance segmentation is a high-precision method in computer for detecting individual animals interest. This can be used in-depth analysis animals, such as examining their subtle interactive behaviors, from videos images. However, existing deep-learning-based instance methods have been mostly developed based on public datasets, which largely omit heavy occlusion problems; therefore, these limitations real-world applications involving object occlusions, farrowing pen systems pig farms crates often impede sow piglets. In this paper, we adapt Center Clustering Network originally designed counting achieve segmentation, dubbed CClusnet-Inseg. Specifically, CClusnet-Inseg uses each pixel predict centers trace form masks clustering results, consists network center offset vector map, Density-Based Spatial Applications with Noise (DBSCAN) algorithm, Centers-to-Mask (C2M), Remain-Centers-to-Mask (RC2M) algorithms. all, 4,600 images were extracted six collected three closed half-open train validate our method. achieves mean average precision (mAP) 84.1 outperforms all other compared study. We conduct comprehensive ablation studies demonstrate advantages effectiveness core modules addition, apply multi-object tracking animal monitoring, predicted that conjunct output could serve an occlusion-resistant representation location object.