sephirex-x / adnet

[ICCV2023] ADNet: Lane Shape Prediction via Anchor Decomposition

ADNet: Lane Shape Prediction via Anchor Decomposition

Pytorch implementation of the paper "ADNet: Lane Shape Prediction via Anchor Decomposition " (ICCV2023 Acceptance)

[paper]

Abstract

In this paper, we revisit the limitations of anchor-based lane detection methods, which have predominantly focused on fixed anchors that stem from the edges of the image, disregarding their versatility and quality. To overcome the inflexibility of anchors, we decompose them into learning the heat map of starting points and their associated directions. This decomposition removes the limitations on the starting point of anchors, making our algorithm adaptable to different lane types in various datasets. To enhance the quality of anchors, we introduce the Large Kernel Attention (LKA) for Feature Pyramid Network (FPN). This significantly increases the receptive field, which is crucial in capturing the sufficient context as lane lines typically run throughout the entire image. We have named our proposed system the Anchor Decomposition Network (ADNet). Additionally, we propose the General Lane IoU (GLIoU) loss, which significantly improves the performance of ADNet in complex scenarios. Experimental results on three widely used lane detection benchmarks, VIL-100, CULane, and TuSimple, demonstrate that our approach outperforms the state-of-the-art methods on VIL-100 and exhibits competitive accuracy on CULane and TuSimple.

Environment setup

1. Clone this repository

git clone https://github.com/Sephirex-X/ADNet.git

2. create conda environment if you using conda

conda create -n ADNet && conda activate ADNet

3. install pytorch and Shapely

conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch

conda install Shapely==1.7.0

4. building independencies

pip install -r requirements.txt

adding -i https://pypi.tuna.tsinghua.edu.cn/simple if you locate in China mainland.

5. setup everything

python setup.py build develop

Benchmarks setup

1. create folder under root path using

mkdir data

2. organize structure like these:

data/
├── CULane -> /mnt/data/xly/CULane/
├── tusimple -> /mnt/data/xly/tusimple/
└── VIL100 -> /mnt/data/xly/VIL100/

under each folder, you may see structure like this:

CULane

Download CULane.

/mnt/data/xly/CULane/
├── driver_100_30frame
├── driver_161_90frame
├── driver_182_30frame
├── driver_193_90frame
├── driver_23_30frame
├── driver_37_30frame
├── laneseg_label_w16
└── list

Tusimple

Download Tusimple.

/mnt/data/xly/tusimple/
├── clips
├── label_data_0313.json
├── label_data_0531.json
├── label_data_0601.json
├── test_label.json
└── test_tasks_0627.json

VIL-100

Download VIL-100.

You may find anno_txt here anno_txt.zip

/mnt/data/xly/VIL100/
├── Annotations
├── anno_txt
├── data
├── JPEGImages
└── Json

Inferencing

1. You can inferencing model using:

python main.py {configs you want to use} --work_dirs {your folder} --load_from {your checkpoint path} --validate --gpus {device id}

for example:

python main.py configs/adnet/tusimple/resnet18_tusimple.py --work_dirs test --load_from best_ckpt/tusimple/res18/best.pth --validate --gpus 3

If you don't assign --work_dirs, it will create folder named work_dirs under root path by default.

2. By adding --view you can see visualization results under folder vis_results, under root path.
3. You can test fps using:

python tools/fps_test.py

Training (only support single gpu)

1. Simply using:

python main.py {configs you want to use} --work_dirs {your folder} --gpus {device id}

2. You can resume from your last checkpoint by:

python main.py {configs you want to use} --work_dirs {your folder} --load_from {your checkpoint path} --gpus {device id}

3. Other hyperparameters can be changed within config files.

4. Your can check training procedure using tensorboard module:

   tensorboard --logdir {your folder} --host=0.0.0.0 --port=1234

Results

VIL-100

Backbone	F1@50	Acc	FP	FN	Download
ResNet-18	89.97	94.23	5.0	5.1	Link
ResNet-34	90.39	94.38	4.4	4.9	Link
ResNet-101	90.90	94.27	4.7	5.0	Link

CULane

Backbone	F1@50	Download
ResNet-18	77.56	Link
ResNet-34	78.94	Link

Tusimple

Backbone	F1@50	Acc	FP	FN	Download
ResNet-18	96.90	96.23	2.91	3.29	Link
ResNet-34	97.31	96.60	2.83	2.53	Link

Citation

If you find our work is useful, please consider citing:

@article{xiao2023adnet,
  title={ADNet: Lane Shape Prediction via Anchor Decomposition},
  author={Xiao, Lingyu and Li, Xiang and Yang, Sen and Yang, Wankou},
  journal={arXiv preprint arXiv:2308.10481},
  year={2023}
}
@InProceedings{Xiao_2023_ICCV,
    author    = {Xiao, Lingyu and Li, Xiang and Yang, Sen and Yang, Wankou},
    title     = {ADNet: Lane Shape Prediction via Anchor Decomposition},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2023},
    pages     = {6404-6413}
}

Acknowledgement

• open-mmlab/mmdetection
• pytorch/vision
• Turoad/lanedet
• Turoad/CLRNet
• lucastabelini/LaneATT
• aliyun/conditional-lane-detection

adnet's Issues

so cool

Missing anno_txt in VIL-100

Hi, Nice Work!
It seems you have an extra annotation fold named anno_txt compare with original dataset.

Can you please provide the generation code? Thanks

drawn backwards ？？？？？

Hello, are the starting positions of shape_hm_gt and hm_gt drawn backwards (the opposite case of symmetry in the middle of the picture)? Because when I look at the original picture, I guess it's upside down

How to export onnx model?

Anyone know how to export onnx model?

Why CULane dataset replication is not successful？

When reproducing the CULane dataset, the evaluation metrics Normal, Crowded, Crowded, Crowded, Crowded, Arrow, Curve, Cross, and Night could not be successfully reproduced

What is python version in this project?

Hi guy, can you tell me what python version you use in this project? Many thanks.

how to get F1 indicators in "Normal", "Crowded" and other scenarios on culane dataset？

Excuse me, I am currently unable to obtain F1 metrics from the culane dataset in "Normal", "Crowded" and other scenarios from the test results, I would appreciate it if you could let me know

Why do the lanes predicted by CLRNet all start from the edges in VIL-100?

Thank you for your excellent work. The visualization results of ADNet are really amazing, but I have a question to ask. I want to know why the anchors of CLRNet are ultimately distributed at the edges of the image. As far as I know, although the anchors of CLRNet, like LineCNN and LaneATT, are pre-defined, their positions and directions can also be learned. Therefore, theoretically, as long as the ground truth does not start from the edges of the image, the lane lines predicted by the trained CLRNet should not start from the edges, right?