/PSANet

PSANet: Point-wise Spatial Attention Network for Scene Parsing, ECCV2018.

Primary LanguageC++

PSANet: Point-wise Spatial Attention Network for Scene Parsing (in construction)

by Hengshuang Zhao*, Yi Zhang*, Shu Liu, Jianping Shi, Chen Change Loy, Dahua Lin, Jiaya Jia, details are in project page.

Introduction

This repository is build for PSANet, which contains source code for PSA module and related evaluation code. For installation, please merge the related layers and follow the description in PSPNet repository (test with CUDA 7.0/7.5 + cuDNN v4).

Note: There are some small structure changes of the evaluation code compared to PSPNet, and has not been fully tested yet for limited time, and PSPNet evaluation should work well if you prefer . After the coming ECCV2018, both PSPNet and PSANet will be updated to CUDA 8.0 + cuDNN v5, both docker file and python interface scripts will be provided, and pytorch implementation will be available in repo semseg later.

Usage

  1. Clone the repository recursively:

    git clone --recursive https://github.com/hszhao/PSANet.git

  2. Merge the caffe layers into PSPNet repository:

    Point-wise spatial attention: pointwise_spatial_attention_layer.hpp/cpp/cu and caffe.proto.

  3. Build Caffe and matcaffe:

    cd $PSANET_ROOT/PSPNet
cp Makefile.config.example Makefile.config
vim Makefile.config
make -j8 && make matcaffe
cd ..

  4. Evaluation:

    • Evaluation code is in folder 'evaluation'.

    • Download trained models and put them in related dataset folder under 'evaluation/model', refer 'README.md'.

    • Modify the related paths in 'eval_all.m':

      Mainly variables 'data_root' and 'eval_list', and your image list for evaluation should be similarity to that in folder 'evaluation/samplelist' if you use this evaluation code structure.

    cd evaluation
vim eval_all.m
    • Run the evaluation scripts:
    ./run.sh

  5. Results:

    Predictions will show in folder 'evaluation/mc_result' and the expected scores are listed as below:

    (mIoU/pAcc. stands for mean IoU and pixel accuracy, 'ss' and 'ms' denote single scale and multiple scale testing.)

    ADE20K:

    network training data testing data mIoU/pAcc.(ss) mIoU/pAcc.(ms) md5sum
    PSANet50 train val 41.92/80.17 42.97/80.92 a8e884
    PSANet101 train val 42.75/80.71 43.77/81.51 ab5e56

    VOC2012:

    network training data testing data mIoU/pAcc.(ss) mIoU/pAcc.(ms) md5sum
    PSANet50 train_aug val 77.24/94.88 78.14/95.12 d5fc37
    PSANet101 train_aug val 78.51/95.18 79.77/95.43 5d8c0f
    PSANet101 COCO + train_aug + val test -/- 85.7/- 3c6a69

    Cityscapes:

    network training data testing data mIoU/pAcc.(ss) mIoU/pAcc.(ms) md5sum
    PSANet50 fine_train fine_val 76.65/95.99 77.79/96.24 25c06a
    PSANet101 fine_train fine_val 77.94/96.10 79.05/96.30 3ac1bf
    PSANet101 fine_train fine_test -/- 78.6/- 3ac1bf
    PSANet101 fine_train + fine_val fine_test -/- 80.1/- 1dfc91

  6. Demo video:

    • Video processed by PSANet (with PSPNet) on BDD dataset for drivable area segmentation: Video.

Citation

If PSANet is useful for your research, please consider citing:

@inproceedings{zhao2018psanet,
  title={{PSANet}: Point-wise Spatial Attention Network for Scene Parsing},
  author={Zhao, Hengshuang and Zhang, Yi and Liu, Shu and Shi, Jianping and Loy, Chen Change and Lin, Dahua and Jia, Jiaya},
  booktitle={ECCV},
  year={2018}
}

Questions

Please contact 'hszhao@cse.cuhk.edu.hk' or 'zy217@ie.cuhk.edu.hk'