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.

Deep Occlusion-Aware Instance Segmentation with Overlapping BiLayers
Lei Ke, Yu-Wing Tai, Chi-Keung Tang
CVPR 2021


  • BCNet: Two/one-stage (detect-then-segment) instance segmentation with state-of-the-art performance.
  • Novelty: A new mask head design, explicit occlusion modeling with bilayer decouple (object boundary and mask) for the occluder and occludee in the same RoI.
  • Efficacy: Large improvements both the FCOS (anchor-free) and Faster R-CNN (anchor-based) detectors.
  • Simple: Small additional computation burden and easy to use.

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 object 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. The heatmap visualization of GCN-1 in fourth column example shows that BCNet handles multiple occluders with in the same RoI by grouping them together. See our paper for more visual examples and comparisons.


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(Two-stage) Backbone Method mAP(mask)
Faster R-CNN Res-R50-FPN Mask R-CNN (ICCV'17) 34.2
Faster R-CNN Res-R50-FPN PANet (CVPR'18) 36.6
Faster R-CNN Res-R50-FPN MS R-CNN (CVPR'19) 35.6
Faster R-CNN Res-R50-FPN PointRend (1x CVPR'20) 36.3
Faster R-CNN Res-R50-FPN BCNet (CVPR'21) 38.4
Faster R-CNN Res-R101-FPN Mask R-CNN (ICCV'17) 36.1
Faster R-CNN Res-R101-FPN MS R-CNN (CVPR'19) 38.3
Faster R-CNN Res-R101-FPN BMask R-CNN (ECCV'20) 37.7
Box-free Res-R101-FPN SOLOv2 (NeurIPS'20) 39.7
Faster R-CNN Res-R101-FPN BCNet (CVPR'21) 39.8
Detector(One-stage) Backbone Method mAP(mask)
FCOS Res-R101-FPN BlendMask (CVPR'20) 38.4
FCOS Res-R101-FPN CenterMask (CVPR'20) 38.3
FCOS Res-R101-FPN SipMask (ECCV'20) 37.8
FCOS Res-R101-FPN CondInst (ECCV'20) 39.1
FCOS Res-R101-FPN BCNet (CVPR'21) 39.6, Pretrained Model, Submission File
FCOS Res-X101 FPN BCNet (CVPR'21) 41.2


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 image 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:
A brief comparison of mask head architectures, see our paper for full details.


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==
# Boundary dependency
pip install scikit-image
# 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.yaml 2>&1 | tee log/train_log.txt

Pretrained Models

FCOS-version download: link

  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_101_FPN.yaml \
	--eval-only MODEL.WEIGHTS ./pretrained_models/xxx.pth 2>&1 | tee log/test_log.txt



Reference script for producing bilayer mask annotation:


The COCO-OCC split:

The COCO-OCC split download: link, which is detailed described in paper.


If you find BCNet useful in your research or refer to the provided baseline results, 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}

Youtube Video | Poster|
Zhihu Reading

Related Work on partially supervised instance segmentation: CPMask