Skip to content

Latest commit

 

History

History
 
 

bezierlanenet

README.md

BézierLaneNet

Rethinking Efficient Lane Detection via Curve Modeling CVPR 2022

Method Overview

State-of-the-art lane detectors are typically based on semantic segmentation (SCNN, RESA) or point detection (LaneATT). However, semantic segmentation requires customized post-processing & cannot deal with a variable number of lanes. Point detection methods are currently anchor-based, with NMS as post-processing. The more natural way would be getting a curve representation directly. Methods like LSTR made the first steps in this direction, but didn't really achieved comparable performance against SOTA methods. BézierLaneNet use a fully convolutional network to predict cubic Bézier curves, the ease of optimization of Bézier control points made it possible for direct curve methods to compete with SOTAs. A fusion of flipped feature maps is also employed to exploit symmetry in the car's front-view. BézierLaneNet (ResNet-34) achieves 75.6 F-1 on CULane, and attained the 1st place (of all published methods) in the LLAMAS leaderboard at its time, while running at 150 FPS in our benchmark.

Results

Training time estimated with single 2080 Ti.

ImageNet pre-training, 3-times average/best.

TuSimple (test)

backbone aug resolution training time precision accuracy (avg) accuracy FP FN
ResNet18 level 1b 360 x 640 5.5h full 95.01% 95.41% 0.0531 0.0458 model | shell
ResNet34 level 1b 360 x 640 6.5h full 95.17% 95.65% 0.0513 0.0386 model | shell

CULane (test)

backbone aug resolution training time precision F1 (avg) F1 normal crowded night no line shadow arrow dazzle
light		 curve crossroad
ResNet18 level 1b 288 x 800 9.9h mix 73.36 73.67 90.22 71.55 68.70 45.30 70.91 84.09 62.49 58.98 996 model | shell
ResNet34 level 1b 288 x 800 11.0h mix 75.30 75.57 91.59 73.20 69.90 48.05 76.74 87.16 69.20 62.45 888 model | shell

LLAMAS (val)

backbone aug resolution training time precision F1 (avg) F1 TP FP FN Precision Recall
ResNet18 level 1b 360 x 640 5.5h mix 95.42 95.52 70515 3102 3520 95.79 95.25 model | shell
ResNet34 level 1b 360 x 640 6.1h mix 96.04 96.11 70959 2667 3076 96.38 95.85 model | shell

Their test performance can be found at the LLAMAS leaderboard.

Profiling

FPS is best trial-avg among 3 trials on a 2080 Ti.

backbone resolution FPS FLOPS(G) Params(M)
ResNet18 360 x 640 212.83 14.77 4.10
ResNet34 360 x 640 149.52 29.85 9.49
ResNet18 288 x 800 210.79 14.66 4.10
ResNet34 288 x 800 144.65 29.54 9.49

Citation

@inproceedings{feng2022rethinking,
  title={Rethinking efficient lane detection via curve modeling},
  author={Feng, Zhengyang and Guo, Shaohua and Tan, Xin and Xu, Ke and Wang, Min and Ma, Lizhuang},
  booktitle={Computer Vision and Pattern Recognition},
  year={2022}
}