Sketch and Refine: Towards Fast and Accurate Lane Detection

---

Install

git clone https://github.com/passerer/SRLane.git
cd SRLane
conda create -n py38 python=3.8 -y # Create a new Python environment, optional.
conda activate py38 
pip install -r requirements.txt

# Install pytorch, modifying the CUDA version accordingly.
pip uninstall urllib3
pip uninstall urllib3
pip install "urllib3>=1.25.4,<1.27"

python setup.py develop

conda install -c conda-forge geos shapely
conda install -c conda-forge pathspec
conda install -c conda-forge tensorboard

imgaug/augmenters/color.py中2487行后加入
        if cls._LUT_CACHE is None:
            cls._LUT_CACHE = cls._generate_lut_table()

DATASET

Download CULane. Then modify dataset_path in configs/datasets/culane.py accordingly.

Train

Here is an example

CUDA_VISIBLE_DEVICES=0 python tools/main.py configs/exp_srlane_culane.py

Test

Performance

Here is an example:

CUDA_VISIBLE_DEVICES=0 python tools/main.py configs/exp_srlane_culane.py --load_from checkpoint/baseline.pth --validate

The results should be:

SET	F1	SET	F1
total	0.7973	noline	0.5565
normal	0.9352	arrow	0.8950
crowd	0.7858	curve	0.7527
hlight	0.7413	cross	1412 (FP)
shadow	0.8190	night	0.7458

Runtime

Here is an example:

CUDA_VISIBLE_DEVICES=0 python tools/analysis_tools/speed_measure.py configs/exp_srlane_culane.py

Acknowledgement

• Turoad/CLRNet
• hirotomusiker/CLRerNet