Clone:
git clone https://github.com/MarvinTeichmann/KittiSeg.git
Initialize all submodules:
git submodule update --init --recursive
We have wrote a convert.launch to transfer rosbag to series of *.jpg with help of Qinghai and got 7507 gray photos from HKUST to Kengkou with a Panoramic camera. Please see ftp for detail.
I picked about 60 images every hundred to mark label in labelme. The labelme will cover road as red and background as black automatically, so don�t forget to change corresponding RGB value in KittiSeg/hypes/KittiSeg.json. The default is road_color: [255,0,255] and background: [255,0,0] and change them into road_color: [255,0,0] and background: [0,0,0]. Otherwise, you may get totally black photo after 5 hours training.
The mask images is in KittiSeg/DATA/data_road/training_DIY/Annotation
Change image channel from 1 to 3 with data_trans.m.
Create path list train3.txt and val3.txt for inputs images which should be like below: 
Tips�Firstly, you'd better write relative path not absolute path as the author did. You can also choose to change the os.path in kitti_seg_inout.py. Secondly, don�t forget the space in original image path and label image path, otherwise, you will get error �need more than 1 value to unpack�.
Change name of val3.txt and train3.txt in hypes/KittiSeg.json into yours.
python train.py --hypes hypes/KittiSeg.json
Then you will get a folder in in RUNS like KittiSeg_2017_12_12_08.30, which contains images folder, several events, model_files and output.log.
Check the road segmentation in images and you can adjust the testing images folder and continue to use this model if it's not bad.
python continue.py - -logdir RUNS/KittiSeg_2017_12_12_08.30
There are three parts that need to change in test.py test.py if you want to run video segmentation, include the videos you have saved or live video. I have marked the three parts.
python
At least 8G memory is needed and you can change in KittiSeg/incl/tensorvision/train.py shown as below:
0.9 means 90% of your whole memory.