Our paper is now available!

You can access it here.😊

@ARTICLE{10954269,
  author={Liu, Binhui and Ling, Qiang},
  journal={IEEE Transactions on Intelligent Transportation Systems}, 
  title={Sparse LaneFormer: End-to-End Lane Detection With Sparse Proposals and Interactions}, 
  year={2025},
  volume={},
  number={},
  pages={1-14},
  keywords={Lane detection;Feature extraction;Proposals;Decoding;Pedestrians;Object detection;Accuracy;Transformers;Polynomials;Annotations;End-to-end lane detection;sparse design;lane aware query;transformer;deformable attention},
  doi={10.1109/TITS.2025.3555220}}

Sparse LaneFormer

This repo is the PyTorch implementation for Sparse LaneFormer. We propose a fully sparse end-to-end lane detection framework called Sparse LaneFormer, which incorporates both sparse proposals and interactions. Specifically, we introduce a Lane Aware Query (LAQ) as sparse lane proposals, which can predict learnable and explicit lane positions. Additionally, we introduce a Lane Token Selector (LTS) and a Linear Deformable Sampler (LDS) to establish sparse interactions among lane tokens. The LTS selectively updates sparse lane tokens in the encoder, while in the decoder, LDS extracts discriminative lane representations from global feature maps by sampling lane features along the LAQ. By focusing on the most informative tokens, Sparse LaneFormer achieves state-of-the-art performance with only 10% of encoder tokens, resulting in an 80% reduction in overall computational cost compared to its dense counterpart.

Installation

1. Clone this repo

git clone https://github.com/ustclbh/Sparse-LaneFormer.git
cd Sparse-LaneFormer

2. Install Pytorch and torchvision

conda create -n sparself python=3.7 -y
conda activate  sparself 
conda install pytorch==1.12.1 torchvision==0.7.0 cudatoolkit=10.2 -c pytorch -y

3. Install other needed packages

pip install -r requirement.txt
# Note: Only tested with mmdet=2.28.2,mmcv=1.7.1.  If you meet errors, please refer to mmdetection repo for more details

4. Fix errors caused by PReLU in MMCV

vim /path/mmcv/cnn/bricks/transformer.py

Then, following pull request in MMCV to solve this problem.

5. Compiling CUDA operators for deformable attention

cd dnlane/models/ops
python setup.py build install
# unit test (should see all checking is True)
python test.py
cd ../../..

Dataset

Download datasets and put it into [data-path] folder. And edit the data_root in the config file to your dataset path.

CULane

[CULane website] [Download]

The directory structure should be like follows:

[data-path]/culane
├── laneseg_label_w16
├── driver_23_30frame
├── driver_37_30frame
├── driver_100_30frame
├── driver_161_90frame
├── driver_182_30frame
├── driver_193_90frame
└── list
    └── test_split
    |   ├── test0_normal.txt
    |   ├── test1_crowd.txt
    |   ├── test2_hlight.txt
    |   ├── test3_shadow.txt
    |   ├── test4_noline.txt
    |   ├── test5_arrow.txt
    |   ├── test6_curve.txt
    |   ├── test7_cross.txt
    |   └── test8_night.txt
    └── train.txt
    └── test.txt
    └── val.txt

TuSimple

[TuSimple website] [Download]

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

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

The directory structure should be like follows:

[data-path]/tusimple
├── clips
├── label_data_0313.json
├── label_data_0531.json
├── label_data_0601.json
├── label_data_0601.json
├── test_label.json
├── test_baseline.json
└── seg_label

Evaluation

CULane

Backbone	F1	Weights	Config	Training log
SwinBase	79.6	sparself_culane_swinb.pth	culane_swinb.py	log

To evalute the model, download the corresponding weights and configs file and run the following commands.

bash eval.sh  /path/to/your/config /path/to/your/checkpoint

Training

if you have multiple gpus:
CUDA_VISIBLE_DEVICES=0,1,2,3   bash dist_train.sh  /path/to/your/config 4

otherwise:
bash run.sh /path/to/your/config

   # Note: We set training epoch to 50 and use an early stop at 30th epoch.

Acknowledgement

• Our code is constructed based on O2SFormer, CLRNet and DAB-DETR. Thanks for their excellent work!