MCTformer (CVPR2022)

Multi-class Token Transformer for Weakly Supervised Semantic Segmentation.

[Paper] [Project Page]

Fig.1 - Overview of MCTformer

🚩 Updates

2023-08-08: MCTformer+ on Arxiv

Environment Setup

• Ubuntu 18.04, with Python 3.6 and the following python dependencies.

pip install -r requirements.txt

Data Preparation

PASCAL VOC 2012

• Download the PASCAL VOC 2012 development kit.

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

• Download augmented annoations SegmentationClassAug.zip from SBD dataset via this link.

• Make your data directory like this below

  VOCdevkit/
  └── VOC2012
      ├── Annotations
      ├── ImageSets
      ├── JPEGImages
      ├── SegmentationClass
      ├── SegmentationClassAug
      └── SegmentationObject

MS COCO 2014

• Download MS COCO 2014 dataset

  wget http://images.cocodataset.org/zips/train2014.zip
  wget http://images.cocodataset.org/zips/val2014.zip

Usage

Train MCTformer+

bash run_mct_plus.sh

Step 1: Run the run.sh script for training MCTformer, visualizing and evaluating the generated class-specific localization maps.

bash run.sh

PASCAL VOC 2012 dataset

Model	Backbone	Google drive
MCTformer-V1	DeiT-small	Weights
MCTformer-V2	DeiT-small	Weights

Step 2: Run the run_psa.sh script for using PSA to post-process the seeds (i.e., class-specific localization maps) to generate pseudo ground-truth segmentation masks. To train PSA, the pre-trained classification weights were used for initialization.

bash run_psa.sh

Step 3: For the segmentation part, run the run_seg.sh script for training and testing the segmentation model. When training on VOC, the model was initialized with the pre-trained classification weights on VOC.

bash run_seg.sh

MS COCO 2014 dataset

Run run_coco.sh for training MCTformer and generating class-specific localization maps. The class label numpy file can be download here. The trained MCTformer-V2 model is here.

bash run_coco.sh

Contact

If you have any questions, you can either create issues or contact me by email lian.xu@uwa.edu.au

Citation

Please consider citing our paper if the code is helpful in your research and development.

@inproceedings{xu2022multi,
  title={Multi-class Token Transformer for Weakly Supervised Semantic Segmentation},
  author={Xu, Lian and Ouyang, Wanli and Bennamoun, Mohammed and Boussaid, Farid and Xu, Dan},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={4310--4319},
  year={2022}
}