Codes for the CNN-based lane detection models. (expected release date: 2018.11.20)
I will release the torch code in 2018.11.20. Tensorflow and pytorch version of the model will be released thereafter.
Preliminary results on CULane testing set (F1-measure):
Category | SCNN-Torch | SCNN-Tensorflow | SCNN-Pytorch |
---|---|---|---|
Normal | 90.6 | 80.0 | -- |
Crowded | 69.7 | 57.1 | -- |
Night | 66.1 | 50.8 | -- |
No line | 43.4 | 35.0 | -- |
Shadow | 66.9 | 45.6 | -- |
Arrow | 84.1 | 68.3 | -- |
Dazzle light | 58.5 | 48.7 | -- |
Curve | 64.4 | 51.6 | -- |
Crossroad | 1990 | 4480 | -- |
Total | 71.6 | 58.7 | -- |
Progress (on-going, debug):
- Define network architecture
- Load pre-trained weights
- Define dataloader (data augmentation [on-going])
- Testing
- Training
- validation
- using multiple GPUs
- clean the codes and make them reproducible
Observations:
FP is big. The output of the lane existence prediction branch is always (1, 1, 1, 1)
.
Notes:
Please go to the Tensorflow-SCNN repo to see detailed instructions.
Progress (on-going):
- Define network architecture (VGG-16 + message passing)
- Load pre-trained weights
- Define dataloader (load images and labels + data augmentation)
- Testing (generate probability maps + smoothing)
- Training
- validation
- using multiple GPUs
- clean the codes and make them reproducible
The ground-truth labels of TuSimple testing set is now available at TuSimple. Please evaluate your pred.json using the labels and this script.
The whole dataset is available at CULane.
If you use the codes, please cite the following publications:
@inproceedings{pan2018SCNN,
author = {Xingang Pan, Jianping Shi, Ping Luo, Xiaogang Wang, and Xiaoou Tang},
title = {Spatial As Deep: Spatial CNN for Traffic Scene Understanding},
booktitle = {AAAI Conference on Artificial Intelligence (AAAI)},
month = {February},
year = {2018}
}
Our paper working on lane detection will be available soon!
This repo is built upon SCNN and LaneNet
If you have any problems in reproducing the results, just raise an issue in this repo.