Deep Occlusion-Aware Instance Segmentation with Overlapping BiLayers

Segmenting highly-overlapping objects is challenging, because typically nodistinction is made between real object contours and occlusion boundaries.Unlike previous two-stage instance segmentation methods, we model imageformation as composition of two overlapping layers, and propose BilayerConvolutional Network (BCNet), where the top GCN layer detects the occludingobjects (occluder) and the bottom GCN layer infers partially occluded instance(occludee). The explicit modeling of occlusion relationship with bilayerstructure naturally decouples the boundaries of both the occluding and occludedinstances, and considers the interaction between them during mask regression.We validate the efficacy of bilayer decoupling on both one-stage and two-stageobject detectors with different backbones and network layer choices. Despiteits simplicity, extensive experiments on COCO and KINS show that ourocclusion-aware BCNet achieves large and consistent performance gain especiallyfor heavy occlusion cases. Code is available at https://github.com/lkeab/BCNet.