hszhao/PSANet

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).

PyTorch Version

Highly optimized PyTorch codebases available for semantic segmentation in repo: semseg, including full training and testing codes for PSPNet and PSANet.

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 '[email protected]' or '[email protected]'