Mask Transfiner for High-Quality Instance Segmentation [Mask Transfiner, CVPR 2022].
This is the official pytorch implementation of Transfiner built on the open-source detectron2. Our project website contains more information, including the visual slider comparison: vis.xyz/pub/transfiner.
Mask Transfiner for High-Quality Instance Segmentation
Lei Ke, Martin Danelljan, Xia Li, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu
CVPR, 2022
๐ฅ๐ฅ We released the Video Mask Transfiner and HQ-YTVIS benchmark in ECCV'2022.
- Transfiner: High-quality instance segmentation with state-of-the-art performance and extreme details.
- Novelty: An efficient transformer targeting for high-resolution instance masks predictions based on the quadtree structure.
- Efficacy: Large mask and boundary AP improvements on three instance segmentation benchmarks, including COCO, Cityscapes and BDD100k.
- Simple: Small additional computation burden compared to standard transformer and easy to use.
- ๐ฅ๐ฅ Play with our Mask Transfiner demo at
, supported by Huggingface Spaces.
(Check Table 9 of the paper for full results, all methods are trained on COCO train2017. This is a reimplementation. Thus, the numbers might be slightly different from the ones reported in our original paper.)
Backbone(configs) | Method | mAP(mask) |
---|---|---|
R50-FPN | Mask R-CNN (ICCV'17) | 34.2 |
R50-FPN | PANet (CVPR'18) | 36.6 |
R50-FPN | MS R-CNN (CVPR'19) | 35.6 |
R50-FPN | PointRend (1x, CVPR'20) | 36.3 |
R50-FPN | Transfiner (1x, CVPR'22) | 37.0, Pretrained Model |
Res-R50-FPN | BCNet (CVPR'21) | 38.4 |
R50-FPN | Transfiner (3x, CVPR'22) | 39.2, Pretrained Model |
R50-FPN-DCN | Transfiner (3x, CVPR'22) | 40.5, Pretrained Model |
Backbone(configs) | Method | mAP(mask) |
---|---|---|
R101-FPN | Mask R-CNN (ICCV'17) | 36.1 |
R101-FPN | MS R-CNN (CVPR'19) | 38.3 |
R101-FPN | BMask R-CNN (ECCV'20) | 37.7 |
R101-FPN | SOLOv2 (NeurIPS'20) | 39.7 |
R101-FPN | BCNet (CVPR'21) | 39.8 |
R101-FPN | Transfiner (3x, CVPR'22) | 40.5, Pretrained Model |
R101-FPN-DCN | Transfiner (3x, CVPR'22) | 42.2, Pretrained Model |
Backbone(configs) | Pretrain | Lr Schd | Size | Method | mAP(box) on Val2017 | mAP(mask) on Val2017 |
---|---|---|---|---|---|---|
Swin-T,init_weight of imagenet (d2 format) | ImageNet-1k | 3X | [480-800] | Transfiner | 46.9 | 43.5, Pretrained Model |
Swin-B,init_weight of imagenet (d2 format) | ImageNet-22k | 3X | [480-800] | Transfiner | 49.8 | 45.5,Pretrained Model |
Backbone(configs) | Lr Schd | Method | mAP(mask) |
---|---|---|---|
X101-FPN | 1x | Mask R-CNN | 27.1 |
X101-FPN | 1x | Transfiner | 29.2, Pretrained Model |
Two-stage and query-based instance segmentation methods have achieved remarkable results. However, their segmented masks are still very coarse. In this paper, we present Mask Transfiner for high-quality and efficient instance segmentation. Instead of operating on regular dense tensors, our Mask Transfiner decomposes and represents the image regions as a quadtree. Our transformer-based approach only processes detected error-prone tree nodes and self-corrects their errors in parallel. While these sparse pixels only constitute a small proportion of the total number, they are critical to the final mask quality. This allows Mask Transfiner to predict highly accurate instance masks, at a low computational cost. Extensive experiments demonstrate that Mask Transfiner outperforms current instance segmentation methods on three popular benchmarks, significantly improving both two-stage and query-based frameworks by a large margin of +3.0 mask AP on COCO and BDD100K, and +6.6 boundary AP on Cityscapes.
conda create -n transfiner python=3.7 -y
conda activate transfiner
conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=11.0 -c pytorch
# 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
pip install kornia==0.5.11
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 transfiner
cd $INSTALL_DIR
git clone --recursive https://github.com/SysCV/transfiner.git
cd transfiner/
python3 setup.py build develop
unset INSTALL_DIR
Prepare for coco2017 dataset and Cityscapes following this instruction.
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
Refer to our scripts folder for more traning, testing and visualization commands:
bash scripts/train_transfiner_3x_101.sh
Or
bash scripts/train_transfiner_1x_50.sh
Download the pretrained models from the above result table:
mkdir pretrained_model
#And put the downloaded pretrained models in this directory.
bash scripts/test_3x_transfiner_101.sh
bash scripts/visual.sh
for swin-based model:
bash scripts/visual_swinb.sh
If you find Mask Transfiner useful in your research or refer to the provided baseline results, please star โญ this repository and consider citing ๐:
@inproceedings{transfiner,
author={Ke, Lei and Danelljan, Martin and Li, Xia and Tai, Yu-Wing and Tang, Chi-Keung and Yu, Fisher},
title={Mask Transfiner for High-Quality Instance Segmentation},
booktitle = {CVPR},
year = {2022}
}
If you are interested in Video Mask Transfiner and High-Quality Video Instance Segmentation data:
@inproceedings{vmt,
title = {Video Mask Transfiner for High-Quality Video Instance Segmentation},
author = {Ke, Lei and Ding, Henghui and Danelljan, Martin and Tai, Yu-Wing and Tang, Chi-Keung and Yu, Fisher},
booktitle = {European Conference on Computer Vision (ECCV)},
year = {2022}
}
Related NeurIPS 2021 Work on multiple object tracking & segmentation: PCAN
Related CVPR 2021 Work on occlusion-aware instance segmentation: BCNet
Related ECCV 2020 Work on partially supervised instance segmentation: CPMask