Awesome

CoSA                                       

Weakly Supervised Co-training with Swapping Assignments for Semantic Segmentation

Xinyu Yang, Hossein Rahmani, Sue Black, Bryan M. Williams

Overview

We propose an end-to-end framework for WSSS: Co-training with Swapping Assignments (CoSA),

Usage

1. Data Preparation

1. Download and Extract COCO 2014

mkdir coco
cd coco
wget http://images.cocodataset.org/zips/train2014.zip
wget http://images.cocodataset.org/zips/val2014.zip
wget http://images.cocodataset.org/zips/test2014.zip
unzip ./train2014.zip
unzip ./val2014.zip
unzip ./test2014.zip

2. Download Segmentation Labels

Here to download the COCO segmentation labels coco_anno.tar and move it to coco dir. Or use the following command to download it directly to the server:

wget https://github.com/youshyee/CoSA/releases/download/ann_coco/coco_anno.tar

After that you should extract it by running:

tar -xvf coco_anno.tar

then you should have a directory structure like this (the number in the brackets is the number of images):

coco/
├── SegmentationClass
│   ├── train2014 (82081)
│   └── val2014 (40137)
├── test2014 (40775)
├── train2014 (82783)
└── val2014 (40504)

1. Download PASCAL VOC 2012 and Extract

wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar
tar –xvf VOCtrainval_11-May-2012.tar

2. Download the augmented annotations

Here is a download link of the augmented annotations. Or use the following command to download it directly to the server:

wget https://github.com/youshyee/CoSA/releases/download/ann_voc/SegmentationClassAug.zip

After downloading SegmentationClassAug.zip, you should unzip it and move it to VOCdevkit/VOC2012. The directory structure should be like this (the number in the brackets is the number of images):

VOCdevkit/
└── VOC2012
    ├── Annotations
    ├── ImageSets
    ├── JPEGImages (17125)
    ├── SegmentationClass
    ├── SegmentationClassAug (12031)
    └── SegmentationObject (2913)

2. Setup Python Environment

We recommend using Anaconda to create a virtual environment.

 conda create -yn cosa python=3.10 pip wheel
 conda activate cosa
 pip install -r requirements.txt

after that, you can install some extension packages: mmcv, bilateralfilter and pydensecrf by running:

mim install mmcv-lite
pip install git+https://github.com/lucasb-eyer/pydensecrf.git
cd utils/bilateralfilter
#sudo apt install swig
swig -python -c++ bilateralfilter.i
python setup.py install

3. Train and Evaluate

### train and eval coco, you may need to modify the `coco_root` in `run_coco.sh` to the path of your COCO dataset.
sh run_coco.sh

### train and eval voc, you may need to modify the `voc12_root` in `run_voc.sh` to the path of your VOC dataset.
sh run_voc.sh

Tested Environment

• Ubuntu 20.04 LTS x86_64
• CUDA 12.1
• NVIDIA GeForce RTX 3090 x2
• Python 3.10

Our Results

Semantic performance on VOC and COCO. Logs and weights are available now.

Visualization results for CoSA comparing with MCT, ToCo and BECO on COCO:

Visualization results for CoSA comparing with MCT, ToCo and BECO on VOC:

The code and weights for CoSA-MS are coming soon.

Citation

Please cite our work if you find it helpful:

@article{yang2024weakly,
  title={Weakly supervised co-training with swapping assignments for semantic segmentation},
  author={Yang, Xinyu and Rahmani, Hossein and Black, Sue and Williams, Bryan M},
  journal={arXiv preprint arXiv:2402.17891},
  year={2024}
}

Acknowledgement

This repo is heavily built upon ToCo and MCT. Please consider citing their works if you find this repo helpful.