Awesome

Spatial CNN for Traffic Lane Detection

Paper

Xingang Pan, Jianping Shi, Ping Luo, Xiaogang Wang, Xiaoou Tang. "Spatial As Deep: Spatial CNN for Traffic Scene Understanding", AAAI2018

This code is modified from fb.resnet.torch.

Introduction

Demo video is available here.

Spatial CNN enables explicit and effective spatial information propagation between neurons in the same layer of a CNN.
It is extremly effective in cases where objects have strong shape priors like the long thin continuous property of lane lines.
VGG16+SCNN outperforms ResNet101 on lane detection.

Requirements

Torch, please follow the installation instructions at fb.resnet.torch.
Matlab (for tools/prob2lines), version R2014a or later.
Opencv (for tools/lane_evaluation), version 2.4.8 (later 2.4.x should also work).
Hardware: For testing, GPU with 3G memory suffices. For training, we recommend 4xGPU with 12G memory.

Before Start

Clone the SCNN repository
```
git clone https://github.com/XingangPan/SCNN.git
```
We'll call the directory that you cloned SCNN as $SCNN_ROOT
Download CULane dataset
```
mkdir -p data/CULane
cd data/CULane
```
Download CULane dataset and extract here. (Note: If you have downloaded the dataset before 16th April 2018, please update the raw annotations of train&val set as described in the dataset website.)
You should have structure like this:
```
$SCNN_ROOT/data/CULane/driver_xx_xxframe    # data folders x6
$SCNN_ROOT/data/CULane/laneseg_label_w16    # lane segmentation labels
$SCNN_ROOT/data/CULane/list                 # data lists
```

Testing

Download our pre-trained models to ./experiments/pretrained
```
cd $SCNN_ROOT/experiments/pretrained
```
Download our best performed model here.
Run test script
```
cd $SCNN_ROOT
sh ./experiments/test.sh
```
Testing results (probability map of lane markings) are saved in experiments/predicts/ by default.
Get curve line from probability map
```
cd tools/prob2lines
matlab -nodisplay -r "main;exit"  # or you may simply run main.m from matlab interface
```
The generated line coordinates would be saved in tools/prob2lines/output/ by default.
Calculate precision, recall, and F-measure
```
cd $SCNN_ROOT/tools/lane_evaluation
make
sh Run.sh   # it may take over 30min to evaluate
```
Note: Run.sh evaluate each scenario separately while run.sh evaluate the whole. You may use calTotal.m to calculate overall performance from all senarios.
By now, you should be able to reproduce our result in the paper.

Training

Download VGG16 pretrained on ImageNet
```
cd $SCNN_ROOT/experiments/models
```
Download VGG16 model here and move it to $SCNN_ROOT/experiments/models/vgg.
Generate SCNN model
```
th SCNN-gen.lua
```
The generated model will be saved in ./vgg_SCNN_DULR9_w9 by default.
Training SCNN
```
cd $SCNN_ROOT
sh ./experiments/train.sh
```
The training process should start and trained models would be saved in $SCNN_ROOT/experiments/models/vgg_SCNN_DULR_w9 by default.
Then you can test the trained model following the Testing steps above. If your model position or name is changed, remember to set them to yours accordingly.

Other Implementations

Tensorflow implementation reproduced by cardwing: https://github.com/cardwing/Codes-for-Lane-Detection.
[new!] Pytorch implementation reproduced by voldemortX: https://github.com/voldemortX/pytorch-auto-drive.

Citing SCNN or CULane

@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}  
}

Acknowledgment

Most work for building CULane dataset is done by Xiaohang Zhan, Jun Li, and Xudong Cao. We thank them for their helpful contribution.