BCNet is a two-stage instance segmentation framework built on detectron2 that explicitly models occlusion by decoupling overlapping objects into bilayer representations. The method improves both anchor-free (FCOS) and anchor-based (Faster R-CNN) detectors by separating object boundary and mask prediction for occluder and occludee within the same RoI. It achieves state-of-the-art results on COCO and KINS benchmarks.