V-LaneNet

V-LaneNet achieves an F1 score of 81.17 on CULane while running at 71 FPS, offering a practical solution for robust lane detection in the wild.

---

Getting Started

Dataset Preparation (CULane)

• Download CULane from the official site: https://xingangpan.github.io/projects/CULane.html
• Expected structure:

data/
  CULane/
    driver_100_30frame/
    driver_161_90frame/
    ...
    list/
      test.txt
      train.txt
      val.txt

• Update data_root in configs/olnet_culane.yaml accordingly.

V-LaneAug Data Generation

V-LaneAug requires SAM and LaMa:

• SAM weights and code: https://github.com/facebookresearch/segment-anything
• LaMa inpainting: https://github.com/advimman/lama

We provide a script to generate occlusion/missing-mark variants:

---

---

Model Zoo

Model	Dataset	F1	FPS	Checkpoint
V-LaneNet (V-LaneAug + V-LaneMixer + V-LaneIoU)	CULane	81.17	71	coming soon

We will release more backbones and training recipes.

---

Roadmap

• Release checkpoints and occlusion data
• Add TensorRT export and INT8 quantization
• Support multilingual logs and docs
• Provide Docker image

---

Acknowledgments

• CULane dataset: Pan et al.
• SAM: Kirillov et al., “Segment Anything”
• LaMa: Suvorov et al., “LaMa: Resolution-robust Large Mask Inpainting with Fourier Convolutions”

We thank the authors and the open-source community for their contributions.

---

Contact

• Maintainer: luzhaoxuan@smail.fjut.edu.cn
• Issues and feature requests: please open a GitHub Issue