This repository contains the PyTorch implementation of:

Intra-class Feature Variation Distillation for Semantic Segmentation, ECCV 2020 [Paper]

All the codes are tested in the following environment:

• Linux (tested on Ubuntu 16.04 / CentOS 7.6)
• Python 3.6.2
• PyTorch 0.4.1
• Single TITAN Xp GPU

• Install PyTorch: conda install pytorch=0.4.1 cuda90 torchvision -c pytorch
• Install other dependences: pip install opencv-python scipy
• Install InPlace-ABN:

cd libs
sh build.sh
python build.py

The build.sh script assumes that the nvcc compiler is available in the current system search path. The CUDA kernels are compiled for sm_50, sm_52 and sm_61 by default. To change this (e.g. if you are using a Kepler GPU), please edit the CUDA_GENCODE variable in build.sh.

• Dataset: [Cityscapes]

• Teacher: PSPNet (ResNet-101) trained on Cityscapes [Google Drive]

• Student: ResNet-18 pretrained on ImageNet [Google Drive]

• After distillation: PSPNet (ResNet-18) [Google Drive]

Please create a new folder ckpt and move all downloaded models to it.

python train.py --data-dir /path/to/cityscapes --save-name /path/to/save --gpu /device/id

python val.py --data-dir /path/to/cityscapes --restore-from /path/to/pth --gpu /device/id

python test.py --data-dir /path/to/cityscapes --restore-from /path/to/pth --gpu /device/id

Please consider citing this work if it helps your research:

@inproceedings{wang2020ifvd,
  title={Intra-class Feature Variation Distillation for Semantic Segmentation},
  author={Wang, Yukang and Zhou, Wei and Jiang, Tao and Bai, Xiang and Xu, Yongchao},
  booktitle={Proceedings of the European Conference on Computer Vision},
  year={2020}
}

This codebase is heavily borrowed from pytorch-segmentation-toolbox and structure_knowledge_distillation. Thanks for their excellent works.