Pytorch implementation of the paper "24-hour Lane Line Detection using via Parallel Scene Information Collaboration" (The paper is under review).

• We construct two generators using unlabeled data to realize the mutual transformation between the daytime scenes and the nighttime scenes. The generated virtual data enhances the diversity of real data, compensates for the limitations of real scenes, and improves the detection accuracy of the model.
• We propose an innovative module, the Multi-Spatial Feature Fusion (MSFF), which preserves the spatial information of both real and generated data. This module enables the joint utilization of real scene and artificial scene data, effectively reducing the impact of artificial noise.
• Experimental results on TuSimple, Night TuSimple, and CULane demonstrate the proposed method's effectiveness for 24-hour lane line detection. This proves the feasibility of applying parallel theory for 24-hour lane line detection.

Only test on Ubuntu 20.04 with:

• Python >= 3.8 (tested with Python3.8.17)
• PyTorch >= 2.0 (tested with Pytorch2.0.1)
• CUDA (tested with cuda_11.3)
• Other dependencies described in requirements.txt

Clone this code to your workspace. We call this directory as $PSIC_ROOT

git clone https://github.com/xiongxiong1996/PSIC

conda create -n PSIC python=3.8 -y
conda activate PSIC

# Install pytorch via pip firstly.
pip install torch torchvision torchaudio

# Install python packages
pip install -r requirements.txt
python setup.py build develop

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

cd $PSIC_ROOT
mkdir -p data
ln -s $CULANEROOT data/CULane

For CULane, you should have structure like this:

$CULANEROOT/data/CULane
	--driver_23_30frame    # data folders x6
	--driver_37_30frame
	--driver_100_30frame
	--driver_161_90frame
	--driver_182_30frame
	--driver_193_90frame
	--laneseg_label_w16    # lane segmentation labels
	--list                 # data lists
                   

Download CULane. Then extract them to $CULANEDAYROOT. We filter CULane's training set, select only the daytime data, modify the $CULANEDAYROOT/list/train_gt.txt, delete the nighttime index. The Modified train_gt.txt can be find on PSIC/tools. Create link to data directory.

cd $PSIC_ROOT
mkdir -p data
ln -s $CULANEDAYROOT data/CULane

For CULane_day, the structure is the same as CULane.

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

cd $PSIC_ROOT
mkdir -p data
ln -s $TUSIMPLEROOT data/tusimple

For Tusimple, you should have structure like this:

$TUSIMPLEROOT
	--clips # data folders
	--lable_data_0313.json # label json file x3
	--lable_data_0531.json 
	--lable_data_0601.json 
	--test_tasks_0627.json
	--test_label.json # test label json file
	--seg_label # run generate_seg_tusimple.py

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

Download Night Tusimple on Baidu Netdisk or Google Drive. Then extract them to $NIGHTTUSIMPLEROOT. Create link to data directory.

cd $PSIC_ROOT
mkdir -p data
ln -s $NIGHTTUSIMPLEROOT data/tusimple_night

Some images of Night Tusimple.

The trained model used in this paper is available day2night_net_G.pth & night2day_net_G.pth. Then move them to $PSIC_ROOT/pth.

In this article, reference to the method described in Arruda et al. 2019, 3,000 images each from daytime and nighttime scenes are used to train the generators.

For training, run

python main.py --config [configs/path_to_your_config] --gpus [gpu_num]

For example, run

python main.py --config configs/PSIC/resnet18_culane.py --gpus 0

For testing, run

python main.py --config [configs/path_to_your_config] --[test|validate] --load_from [path_to_your_model] --gpus [gpu_num]

For example, run

python main.py --config configs/PSIC/resnet18_culane.py --validate --load_from pth/culane_resnet18.pth --gpus 0

Currently, this code can output the visualization result when testing, just add --view. We will get the visualization result in work_dirs/xxx/xxx/visualization.

Visualization results of UFLD, RESA, CLRNet and PSIC on Tusimple and Night Tusimple.

State-of-the-art results on TuSimple. Results was computed using the official source code. Best results are in bold.

State-of-the-art results on Night TuSimple. Results was computed using the official source code. Best results are in bold.

Comparison with state-of-the-art results on CULane dataset with IoU threshold = 0.5. For cross, only FP are shown. Best results are in bold and second best underlined.

If our paper and code are beneficial to your work, please consider citing:

@InProceedings{**,
    author    = {**,
    title     = {**,
    booktitle = {**,
    month     = {**,
    year      = {**,
    pages     = {**
**

• pytorch/vision
• junyanz/CycleGAN
• LCAD-UFES/publications-arruda-ijcnn-2019
• Turoad/CLRNet
• harryhan618/SCNN_Pytorch
• ZJULearning/resa
• cfzd/Ultra-Fast-Lane-Detection
• aliyun/conditional-lane-detection
• SoulmateB/CurveLanes
• RanHao-cq/FLAMNet