HSDF-Lane: Height-Aligned Signed Distance Field with Semantic Lane Prior for 3D Lane Detection

This is the official implementation of HSDF-Lane (ECCV 2026).

---

Dataset Preparation

Please follow data preparation to download dataset.

---

Installation

1. Clone this repository:

git clone https://github.com/JiyongBoo/HSDF-Lane

2. Create a virtual environment:

cd ./HSDF-Lane
conda create -n hsdflane python=3.8.20
pip install -r requirement.txt
conda install -c conda-forge cudatoolkit-dev=11.1

3. Clone the required dependency (Deformable-DETR) in the same parent directory:

cd ..
git clone https://github.com/fundamentalvision/Deformable-DETR.git

Place both repositories under the same parent directory as follows:

<your_workspace>/
├── dataset/
├── Deformable-DETR/
└── HSDF-Lane/

4. Compile CUDA operators:

Build the CUDA operators from the Deformable-DETR ops directory:

export CUDA_HOME=$CONDA_PREFIX
export PATH=$CUDA_HOME/bin:$PATH
export LD_LIBRARY_PATH=$CUDA_HOME/lib:$LD_LIBRARY_PATH

cd ./Deformable-DETR/models/ops
sh ./make.sh

5. Additional Requirements:

pip install setuptools==69.5.1 wheel ninja Cython

---

Eval

Pretrained Checkpoints

Download the pretrained model from the following link: https://huggingface.co/boo0828/HSDF-Lane

cd <your_workspace>/HSDF-Lane/
huggingface-cli download boo0828/HSDF-Lane <ckpt_name> --local-dir ./

Dataset	ckpt_name	Metrics
OpenLane-1000	hsdflane.pth	F1=66.3%
OpenLane-1000 (FPN version)	hsdflane_fpn.pth	F1=66.9%
Apollo-standard,rare	hsdflane_apollo_standard.pth	F1=98.8%
Apollo-illus_chg	hsdflane_apollo_illus_chg.pth	F1=97.9%

Validation

With the dataset and checkpoint in place, run:

cd <your_workspace>/HSDF-Lane/
python tools/val_openlane.py \\
    --config ./tools/hsdflane_config.py \\
    --checkpoint ./hsdflane.pth 

This runs evaluation on the OpenLane validation set.

To evaluate the model on the Apollo dataset, run:

cd <your_workspace>/HSDF-Lane/
python tools/val_apollo.py \\
    --config ./tools/hsdflane_apollo_config.py \\
    --checkpoint ./hsdflane_apollo_standard.pth 

Train

To train HSDF-Lane from scratch:

cd <your_workspace>/HSDF-Lane/
# OpenLane
python tools/train_openlane.py \\
    --config ./tools/hsdflane_config.py
# Apollo
python tools/train_apollo.py \\
    --config ./tools/hsdflane_apollo_config.py

---

Citation

@misc{boo2026hsdflaneheightalignedsigneddistance,
      title={HSDF-Lane: Height-Aligned Signed Distance Field with Semantic Lane Prior for 3D Lane Detection}, 
      author={Jiyong Boo and Byeongin Joung and Hyemin Yang and Kuk-Jin Yoon},
      year={2026},
      eprint={2606.31172},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.31172}, 
}

---

Acknowledgments

This repository builds upon

• HeightLane
• SC-Lane

We are grateful to the authors of these works for providing the foundation for our project.