XingangPan/SCNN

Spatial CNN for Traffic Lane Detection

Paper

Xingang Pan, Xiaohang Zhan, 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

1. Clone the SCNN repository

   git clone https://github.com/XingangPan/SCNN.git
   

   We'll call the directory that you cloned SCNN as $SCNN_ROOT

2. 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

1. Download our pre-trained models to ./experiments/pretrained

   cd $SCNN_ROOT/experiments/pretrained
   

   Download our best performed model here.

2. Run test script

   cd $SCNN_ROOT
   sh ./experiments/test.sh
   

   Testing results (probability map of lane markings) are saved in experiments/predicts/ by default.

3. 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.

4. 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

1. Download VGG16 pretrained on ImageNet

   cd $SCNN_ROOT/experiments/models
   

   Download VGG16 model here and move it to $SCNN_ROOT/experiments/models/vgg.
2. Generate SCNN model

   th SCNN-gen.lua
   

   The generated model will be saved in ./vgg_SCNN_DULR9_w9 by default.
3. 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, Xiaohang Zhan, 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.