Instance Segmentation Based on Mask Scoring R-CNN for Group-housed Pigs

Abstract

Pig instance segmentation is a challenging task for group-housed pigs in complex scenes due to variable light and pig's occlusion. In this paper, an instance segmentation approach based on Mask Scoring R-CNN framework is proposed to solve pig instance segmentation from complex backgrounds automatically. The proposed approach includes three steps: first, a CNN-based feature extraction, which uses backbone architecture with residual network of depth 101 layers (ResNet-101) and feature pyramid network (FPN) that can extract features from low-level and high-level of the feature pyramid according to different scales. Then, the candidate regions of interest (RoIs) are generated by region proposal network (RPN) based on the features derived from backbone architecture. Finally, we detect and segment the pig for each RoI. The proposed approach was trained and tested on dataset from front-view and top-view videos of pigs. The proposed approach achieved best F1 score of 0.9405. According to these experimental results, the approach can robustly detect and segment multiple target pigs under the group-housed pig natural scenes such as uneven illumination, pigs' occlusion and overlapping.