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.
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. |
(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:
 |
 |
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==4.4.0.40
# Boundary dependency
pip install scikit-image
export INSTALL_DIR=$PWD
# install pycocotools. Please make sure you have installed cython.
cd $INSTALL_DIR
git clone https://github.com/cocodataset/cocoapi.git
cd cocoapi/PythonAPI
python setup.py build_ext install
# install BCNet
cd $INSTALL_DIR
git clone https://github.com/lkeab/BCNet.git
cd BCNet/
python3 setup.py build develop
unset INSTALL_DIR
Prepare for coco2017 dataset following this instruction. And use our converted mask annotations to replace original annotation file for bilayer decoupling training.
mkdir -p datasets/coco
ln -s /path_to_coco_dataset/annotations datasets/coco/annotations
ln -s /path_to_coco_dataset/train2017 datasets/coco/train2017
ln -s /path_to_coco_dataset/test2017 datasets/coco/test2017
ln -s /path_to_coco_dataset/val2017 datasets/coco/val2017
bash all.sh
Or
CUDA_VISIBLE_DEVICES=0,1 python3 tools/train_net.py --num-gpus 2 \
--config-file configs/fcos/fcos_imprv_R_50_FPN.yaml 2>&1 | tee log/train_log.txt
FCOS-version download: link
mkdir pretrained_models
#And put the downloaded pretrained models in this directory.
export PYTHONPATH=$PYTHONPATH:`pwd`
CUDA_VISIBLE_DEVICES=0,1 python3 tools/train_net.py --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
bash visualize.sh
bash process.sh
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 📝:
@inproceedings{ke2021bcnet,
author = {Ke, Lei and Tai, Yu-Wing and Tang, Chi-Keung},
title = {Deep Occlusion-Aware Instance Segmentation with Overlapping BiLayers},
booktitle = {CVPR},
year = {2021}
}
Related occlusion handling work:
@inproceedings{ke2021voin,
author = {Ke, Lei and Tai, Yu-Wing and Tang, Chi-Keung},
title = {Occlusion-Aware Video Object Inpainting},
booktitle = {ICCV},
year = {2021}
}
Youtube Video | Poster| Zhihu Reading
Related NeurIPS 2021 Work on multiple object tracking & segmentation: PCAN
Related ECCV 2020 Work on partially supervised instance segmentation: CPMask
BCNet is released under the MIT license. See LICENSE for additional details. Thanks to the Third Party Libs detectron2.
Leave github issues or please contact 'lkeab@cse.ust.hk'