Semi-supervised Semantic Segmentation with Mutual Knowledge Distillation

This project hosts the codo for implementing the MKD algorithm for semi-supervised learning

Installation

Requirements

• Linux (Windows is not officially supported)
• Python 3.6+
• PyTorch 1.10 or higher
• CUDA 9.0 or higher
• GCC 4.9 or higher
• mmcv-full

Install MKD

a. Create a conda virtual environment and activate it.

conda create -n mkd python=3.6 -y
conda activate mkd

b. Install PyTorch and torchvision following the official instructions, e.g.,

conda install pytorch=1.10.0 torchvision cudatoolkit=10.2 -c pytorch -y

c. Install mmcv-full.

pip install mmcv-full==1.4.8 -f https://download.openmmlab.com/mmcv/dist/cu102/torch1.10.0/index.html

d. Install other third-party libraries.

pip install terminaltables imgaug onnxruntime==1.6.0 onnx albumentations Scikit-Image pycocotools tensorboard pillow==8.4.0

e. Clone the MKD repository.

git clone https://github.com/jianlong-yuan/semi-mmseg.git
cd semi-mmseg

f. Install.

pip install -r requirements.txt
pip install -e .  # or "python setup.py develop"

Prepare datasets

It is recommended to symlink your dataset root (assuming YOUR_DATA_ROOT) to $semi-mmseg/data. If your folder structure is different, you may need to change the corresponding paths in config files.

Prepare PASCAL VOC 2012 and Cityscapes

Assuming that you usually store datasets in $YOUR_DATA_ROOT (e.g.,/home/YOUR_NAME/data/).

The different split lists will be store in data directory.

MKD
├── configs
├── data
│   ├── cityscapes
│   │   ├── images
│   │   ├── segmentation
│   |   ├── splits
│   │   |   |    ├── cps_splits
│   │   |   |    ├── u2pl_splits
│   ├── VOCdevkit
│   │   ├── VOC2012
│   │   |   ├── Annotations
│   │   |   ├── ImageSets
│   │   |   ├── JPEGImages
│   │   |   ├── SegmentationClass
│   │   |   ├── SegmentationClassAug
│   │   |   ├── SegmentationObject
│   │   |   ├── splits
│   │   |   |    ├── cps_splits
│   │   |   |    ├── pseudoseg_splits
│   │   |   |    ├── u2pl_splits

Training

./tools/dist_train.sh configs/semi_ablations/cps_meanteacher_3b_w1.5_w.1.0_fdmt1.5.py

Acknowledgement

We would like to thank the MMSegmentation for its open-source project.