Deep Occlusion-Aware Instance Segmentation with Overlapping BiLayers [BCNet, CVPR 2021]

This is the official pytorch implementation of BCNet built on the open-source detectron2.

Lei Ke, Yu-Wing Tai, Chi-Keung Tang
CVPR 2021

  • Two-stage instance segmentation with state-of-the-art performance.
  • Image formation as composition of two overlapping layers.
  • Bilayer decoupling for the occluder and occludee.
  • Efficacy on both the FCOS and Faster R-CNN detectors.

Under construction. Our code and pretrained model will be fully released in two months.

Visualization of Occluded Objects


Qualitative instance segmentation results of our BCNet, using ResNet-101-FPN and Faster R-CNN detector. The bottom row visualizes squared heatmap of contour and mask predictions by the two GCN layers for the occluder and occludee in the same ROI region specified by the red bounding box, which also makes the final segmentation result of BCNet more explainable than previous methods.


Qualitative instance segmentation results of our BCNet, using ResNet-101-FPN and FCOS detector.

Results on COCO test-dev

(Check Table 8 of the paper for full results, all methods are trained on COCO train2017)

Detector Backbone Method mAP(mask)
Faster R-CNN ResNet-50 FPN Mask R-CNN 34.2
Faster R-CNN ResNet-50 FPN MS R-CNN 35.6
Faster R-CNN ResNet-50 FPN PointRend 36.3
Faster R-CNN ResNet-50 FPN PANet 36.6
Faster R-CNN ResNet-50 FPN BCNet 38.4
Faster R-CNN ResNet-101 FPN Mask R-CNN 36.1
Faster R-CNN ResNet-101 FPN BMask R-CNN 37.7
Faster R-CNN ResNet-101 FPN MS R-CNN 38.3
Faster R-CNN ResNet-101 FPN BCNet 39.8, [Pretrained Model]
FCOS ResNet-101 FPN SipMask 37.8
FCOS ResNet-101 FPN BlendMask 38.4
FCOS ResNet-101 FPN CenterMask 38.3
FCOS ResNet-101 FPN BCNet 39.6, [Pretrained Model]


Segmenting highly-overlapping objects is challenging, because typically no distinction is made between real object contours and occlusion boundaries. Unlike previous two-stage instance segmentation methods, BCNet models image formation as composition of two overlapping layers, where the top GCN layer detects the occluding objects (occluder) and the bottom GCN layer infers partially occluded instance (occludee). The explicit modeling of occlusion relationship with bilayer structure naturally decouples the boundaries of both the occluding and occluded instances, and considers the interaction between them during mask regression. We validate the efficacy of bilayer decoupling on both one-stage and two-stage object detectors with different backbones and network layer choices. The network of BCNet is as follows:


Step-by-step Installation

conda create -n bcnet python=3.7 -y
source activate bcnet
conda install pytorch==1.4.0 torchvision==0.5.0 cudatoolkit=10.1 -c pytorch
# FCOS and coco api and visualization dependencies
pip install ninja yacs cython matplotlib tqdm
pip install opencv-python==
# install pycocotools. Please make sure you have installed cython.
git clone
cd cocoapi/PythonAPI
python build_ext install
# install BCNet
git clone
cd BCNet/
python3 build develop

Dataset Preparation

Prepare for coco2017 dataset following this instruction. And use our converted mask annotations to replace original annotation file for bilayer decoupling training.

  mkdir -p datasets/coco
  ln -s /path_to_coco_dataset/annotations datasets/coco/annotations
  ln -s /path_to_coco_dataset/train2017 datasets/coco/train2017
  ln -s /path_to_coco_dataset/test2017 datasets/coco/test2017
  ln -s /path_to_coco_dataset/val2017 datasets/coco/val2017

Multi-GPU Training and evaluation on Validation set



CUDA_VISIBLE_DEVICES=0,1 python3 tools/ --num-gpus 2 \
	--config-file configs/fcos/fcos_imprv_R_50_FPN_1x.yaml 2>&1 | tee log/train_log.txt

Pretrained Models


  mkdir pretrained_models
  #And put the downloaded pretrained models in this directory.

Testing on Test-dev

CUDA_VISIBLE_DEVICES=0,1 python3 tools/ --num-gpus 2 \
	--config-file configs/fcos/fcos_imprv_R_50_FPN_1x.yaml \
	--eval-only MODEL.WEIGHTS ../pretrained_models/xxx.pth 2>&1 | tee log/test_log.txt


Check for installation instructions.
Prepare Data

  mkdir -p datasets/coco
  ln -s /path_to_coco_dataset/annotations datasets/coco/annotations
  ln -s /path_to_coco_dataset/train2014 datasets/coco/train2014
  ln -s /path_to_coco_dataset/test2014 datasets/coco/test2014
  ln -s /path_to_coco_dataset/val2014 datasets/coco/val2014

Pretrained Models

  mkdir pretrained_models
  #The pretrained models will be downloaded when running the program.

My training log and pre-trained models can be found here link or link(pw:xm3f).

Single GPU Training

  python tools/ --config-file "configs/e2e_ms_rcnn_R_50_FPN_1x.yaml" SOLVER.IMS_PER_BATCH 2 SOLVER.BASE_LR 0.0025 SOLVER.MAX_ITER 720000 SOLVER.STEPS "(480000, 640000)" TEST.IMS_PER_BATCH 1

Multi-GPU Training

  export NGPUS=8
  python -m torch.distributed.launch --nproc_per_node=$NGPUS tools/ --config-file "configs/e2e_ms_rcnn_R_50_FPN_1x.yaml" 



If you find BCNet useful in your research, please star this repository and consider citing:

    author = {Ke, Lei and Tai, Yu-Wing and Tang, Chi-Keung},
    title = {Deep Occlusion-Aware Instance Segmentation with Overlapping BiLayers},
    booktitle = {CVPR},
    year = {2021},