RESA

PyTorch implementation of the paper "RESA: Recurrent Feature-Shift Aggregator for Lane Detection".

Our paper has been accepted by AAAI2021.

News: We also release RESA on LaneDet. It's also recommended for you to try LaneDet.

Introduction

RESA shifts sliced feature map recurrently in vertical and horizontal directions and enables each pixel to gather global information.
RESA achieves SOTA results on CULane and Tusimple Dataset.

Get started

Clone the RESA repository
```
git clone https://github.com/zjulearning/resa.git
```
We call this directory as $RESA_ROOT
Create a conda virtual environment and activate it (conda is optional)
```
conda create -n resa python=3.8 -y
conda activate resa
```

Install dependencies

# Install pytorch firstly, the cudatoolkit version should be same in your system. (you can also use pip to install pytorch and torchvision)
conda install pytorch torchvision cudatoolkit=10.1 -c pytorch

# Or you can install via pip
pip install torch torchvision

# Install python packages
pip install -r requirements.txt

Data preparation

Download CULane and Tusimple. Then extract them to $CULANEROOT and $TUSIMPLEROOT. Create link to data directory.

cd $RESA_ROOT
mkdir -p data
ln -s $CULANEROOT data/CULane
ln -s $TUSIMPLEROOT data/tusimple

For CULane, you should have structure like this:

$CULANEROOT/driver_xx_xxframe    # data folders x6
$CULANEROOT/laneseg_label_w16    # lane segmentation labels
$CULANEROOT/list                 # data lists

For Tusimple, you should have structure like this:

$TUSIMPLEROOT/clips # data folders
$TUSIMPLEROOT/lable_data_xxxx.json # label json file x4
$TUSIMPLEROOT/test_tasks_0627.json # test tasks json file
$TUSIMPLEROOT/test_label.json # test label json file

For Tusimple, the segmentation annotation is not provided, hence we need to generate segmentation from the json annotation.

python tools/generate_seg_tusimple.py --root $TUSIMPLEROOT
# this will generate seg_label directory

Install CULane evaluation tools.

This tools requires OpenCV C++. Please follow here to install OpenCV C++. Or just install opencv with command sudo apt-get install libopencv-dev

Then compile the evaluation tool of CULane.
```
cd $RESA_ROOT/runner/evaluator/culane/lane_evaluation
make
cd -
```
Note that, the default opencv version is 3. If you use opencv2, please modify the OPENCV_VERSION := 3 to OPENCV_VERSION := 2 in the Makefile.

Training

For training, run

python main.py [configs/path_to_your_config] --gpus [gpu_ids]

For example, run

python main.py configs/culane.py --gpus 0 1 2 3

Testing

For testing, run

python main.py c[configs/path_to_your_config] --validate --load_from [path_to_your_model] [gpu_num]

For example, run

python main.py configs/culane.py --validate --load_from culane_resnet50.pth --gpus 0 1 2 3

python main.py configs/tusimple.py --validate --load_from tusimple_resnet34.pth --gpus 0 1 2 3

We provide two trained ResNet models on CULane and Tusimple, downloading our best performed model (Tusimple: GoogleDrive/BaiduDrive(code:s5ii), CULane: GoogleDrive/BaiduDrive(code:rlwj) )

Visualization

Just add --view.

For example:

python main.py configs/culane.py --validate --load_from culane_resnet50.pth --gpus 0 1 2 3 --view

You will get the result in the directory: work_dirs/[DATASET]/xxx/vis.

Citation

@misc{zheng2020resa,
      title={RESA: Recurrent Feature-Shift Aggregator for Lane Detection}, 
      author={Tu Zheng and Hao Fang and Yi Zhang and Wenjian Tang and Zheng Yang and Haifeng Liu and Deng Cai},
      year={2020},
      eprint={2008.13719},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}