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, 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

Clone the repository:

git clone git@github.com:TuSimple/TuSimple-DUC.git
python setup.py develop --user

Download the pretrained model from Google Drive.

Build MXNet (only tested on the TuSimple version):

git clone --recursive git@github.com:TuSimple/mxnet.git
vim make/config.mk
make -j
cd python
python setup.py develop --user

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

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.

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.

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 panqu.wang@tusimple.ai or pengfei.chen@tusimple.ai .