【Hackathon + No.161】论文复现：CLRNet: Cross Layer Refinement Network for Lane Detection

[khan-yin]

Environment

• PaddlePaddle2.4.0
• V100
• need pip install imgaug>=0.4.0 for image augmentations in requirements.txt

imgaug>=0.4.0

Training

• single GPU

python tools/train.py -c configs/clrnet/clr_resnet18_culane.yml -o use_gpu=true

• multi GPU

export CUDA_VISIBLE_DEVICES=0,1,2,3
python -m paddle.distributed.launch --gpus 0,1,2,3 tools/train.py -c configs/clrnet/clr_resnet18_culane.yml

P.S. recommend single GPU training for better accuracy.

Evaluation

mkdir evaluation
python tools/eval.py -c configs/clrnet/clr_resnet18_culane.yml -o weights=output/clr_resnet18_culane/model_final.pdparams output_eval=evaluation/ use_gpu=true

Inference

python tools/infer_culane.py -c configs/clrnet/clr_resnet18_culane.yml -o weights=output/clr_resnet18_culane/model_final.pdparams use_gpu=true --infer_img=demo/lane00000.jpg

Weights&Logs

backbone	mF1	F1@50	F1@75	download	code
ResNet-18	54.98	79.46	62.10	BaiduDisk	fz71

.
|-- model_final.pdparams
|-- evaluation.log
|-- training.log

Demo

[paddle-bot]

Thanks for your contribution!