SGSN
Supervised Generative Segmentation Network for road segmentation for all-day outdoor robot navigation.
Major Requirements
Python 2.7.12
Numpy 1.13.1
Tensorflow 1.2.0
cuda 8.0
cudnn 5.1.10
OpenCV 2.4.13
Running codes
1. Build six folders
(YOUR_PATH)/SGSN/datasets/train/X/images/
(YOUR_PATH)/SGSN/datasets/train/X/labels/
(YOUR_PATH)/SGSN/datasets/train/Y/images/
(YOUR_PATH)/SGSN/datasets/train/Y/labels/
(YOUR_PATH)/SGSN/datasets/test/X/images/
(YOUR_PATH)/SGSN/datasets/test/X/labels/
2. Download UAS dataset
Please follow the links in the Download links.
3. Copy images and labels
To build training set, copy the training images and the labels (not label graphs) of dusk_sight, night_sight, rain_sight and sun_sight to X domain (./datasets/train/X/images/, ./datasets/train/X/labels), and copy the images and the labels of dusk_sight to Y domain (./datasets/train/Y/images/, ./datasets/train/Y/labels).
To build test set, copy the test images and the labels (not label graphs) of dusk_sight, night_sight, rain_sight and sun_sight to X domian (./datasets/test/X/images/, ./datasets/test/X/labels).
4. Duplicate images and labels of night in training set
cd (YOUR_PATH)/SGSN/datasets
python2 ./copy_night.py
5. Train SGSN
cd (YOUR_PATH)/SGSN
python2 ./train.py
6. Evaluate SGSN
cd (YOUR_PATH)/SGSN
python2 ./evaluate.py
Citation
UAS (UESTC All-Day Scenery)
Introduction
The dataset for road segmentation for the outdoor images in any time.
A total of 6380 images of different time and weather conditions are contained in UAS, including 1399 samples captured at dusk, 2167 samples captured at night, 819 samples captured in the rain and 1995 samples captured in the sunshine. For each image, a precise binary semantic label is made by manual annotation.
Download links
https://pan.baidu.com/s/1IWSVKYBrYwxaRThPfDsDGg
Citation
@article{zhang2018road,
title={Road Segmentation for All-Day Outdoor Robot Navigation},
author={Zhang, Yuxiao and Chen, Haiqiang and He, Yiran and Ye, Mao and Cai, Xi and Zhang, Dan},
journal={Neurocomputing},
year={2018},
publisher={Elsevier}
}