Skip to content

hustvl/BoxCaseg

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

91 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Weakly-supervised Instance Segmentation via Class-agnostic Learning with Salient Images.(CVPR 2021)

This is the official code of the paper Weakly-supervised Instance Segmentation via Class-agnostic Learning with Salient Images, published in CVPR 2021, by Xinggang Wang, Jiapei Feng, Bin Hu, Qi Ding, Longjin Ran, Xiaoxin Chen, Wenyu Liu.

Table of Contents
  1. Introduction
  2. Prerequisites
  3. Usage
  4. Citation
  5. License
  6. Acknowledgements

Introduction

pipeline

Weakly-supervised instance segmentation uses coarser annotations to acquire a high-precision instance segmentation model, such as bounding boxes. This "box-to-seg" process is a class-agnostic process. Our motivation is to learn a model that achieves generic class-agnostic segmentation.

The training process is divided into three steps. Firstly, we obtain a class-agnostic segmentation model through the joint training of box-supervised dataset and salient object segmentation dataset. Secondly, we use the class-agnostic segmentation model to produce high-quality predictions for training instances. Those segmentation predictions are merged to generate proxy instance masks for training images. Finally, we re-train a Mask R-CNN with proxy masks.

Prerequisites

This is an example of how to list things you need to use the software and how to install them.

  • Python>=3.6, PyTorch
  • Augmented PASCAL VOC 2012 Dataset (10582 images for training)
  • DUT-TR-Single Dataset

Usage

Here, we provide the DUT-TR-Single datasets, the cocostyle annotations of PASCAL VOC, our pre-trained models and proxy masks for training set.

Training

  • After downloading the dataset, put them in the specific folder. Then, cd jointraining and run the following command to do joint training:
    bash train_pd.sh

for evaluation:

    bash pascal_val.sh

for predicting segmentation maps of the training instances:

    bash pascal_psd_mask.sh

More detailed instructions are provided in the jointraining.

  • Generate the proxy masks for box-supervised dataset. Run the following command:
    cd proxy_mask
    python pascal_proxy_mask_generate.py --gt-path training_set_boundingbox_cocostyle_json --seg-pred predicted_results
  • Retrain a Mask R-CNN. We use the Mask R-CNN as our instance segmentation framework and we modify two file include ../maskrcnn_benchmark/data/datasets/coco.py and ../maskrcnn_benchmark/modeling/roi_heads/mask_head/loss.py.

Run the following command:

    cd retrain/maskrcnn-benchmark/
    python -m torch.distributed.launch --nproc_per_node=2 ./tools/train_net.py --config-file e2e_mask_rcnn_R_101_FPN_4x_voc_aug_cocostyle.yaml

Check INSTALL.md for installation instructions.

Citation

If you find the code useful in your research, please consider citing:

@inproceedings{wang2021boxcaseg,
  title     =  {Weakly-supervised Instance Segmentation via Class-agnostic Learning with Salient Images},
  author    =  {Wang, Xinggang and Feng, Jiapei and Hu, Bin and Ding, Qi and Ran, Longjin and Chen, Xiaoxin and Liu, Wenyu},
  booktitle =  {Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR)},
  year      =  {2021}
}

License

Distributed under the MIT License. See LICENSE for more information.

Acknowledgements

  • We thank NSFC, Zhejiang Laboratory and VIVO Inc for their support to this project.
  • The code is borrowed from HRNet, BBTP and Mask R-CNN.

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 3

  •  
  •  
  •