TuSimple-DUC

by Panqu Wang, Pengfei Chen, Ye Yuan, Ding Liu, Zehua Huang, Xiaodi Hou, and Garrison Cottrell.

Introduction

This repository is for Understanding Convolution for Semantic Segmentation (WACV 2018), which achieved state-of-the-art result on the CityScapes, PASCAL VOC 2012, and Kitti Road benchmark.

Requirement

We tested our code on:

Ubuntu 16.04, Python 2.7 with

MXNet (0.11.0), numpy(1.13.1), cv2(3.2.0), PIL(4.2.1), and cython(0.25.2)

Usage

1. Clone the repository:

   git clone [email protected]:TuSimple/TuSimple-DUC.git
   python setup.py develop --user

2. Download the pretrained model from Google Drive.

3. Build MXNet (only tested on the TuSimple version):

   git clone --recursive [email protected]:TuSimple/mxnet.git
   vim make/config.mk (we should have USE_CUDA = 1, modify USE_CUDA_PATH, and have USE_CUDNN = 1 to enable GPU usage.)
   make -j
   cd python
   python setup.py develop --user

   For more MXNet tutorials, please refer to the official documentation.

4. Training:

   cd train
   python train_model.py ../configs/train/train_cityscapes.cfg

   The paths/dirs in the .cfg file need to be specified by the user.

5. Testing

   cd test
   python predict_full_image.py ../configs/test/test_full_image.cfg
   

   The paths/dirs in the .cfg file need to be specified by the user.

6. Results:

   Modify the result_dir path in the config file to save the label map and visualizations. The expected scores are:

   (single scale testing denotes as 'ss' and multiple scale testing denotes as 'ms')

   • ResNet101-DUC-HDC on CityScapes testset (mIoU): 79.1(ss) / 80.1(ms)
   • ResNet152-DUC on VOC2012 (mIoU): 83.1(ss)

Citation

If you find the repository is useful for your research, please consider citing:

@article{wang2017understanding,
  title={Understanding convolution for semantic segmentation},
  author={Wang, Panqu and Chen, Pengfei and Yuan, Ye and Liu, Ding and Huang, Zehua and Hou, Xiaodi and Cottrell, Garrison},
  journal={arXiv preprint arXiv:1702.08502},
  year={2017}
}

Questions

Please contact [email protected] or [email protected] .