Awesome
CLRerNet Official Implementation
The official implementation of our paper "CLRerNet: Improving Confidence of Lane Detection with LaneIoU", by Hiroto Honda and Yusuke Uchida.
What's New
- Code for training is available ! (Dec. 1, 2023)
- Our CLRerNet paper has been accepted to WACV2024 ! (Oct. 25, 2023)
- LaneIoU loss and cost are published. (PR#17, Oct.22, 2023)
Method
CLRerNet features LaneIoU for the target assignment cost and loss functions aiming at the improved quality of confidence scores.<br> LaneIoU takes the local lane angles into consideration to better correlate with the segmentation-based IoU metric.
<p align="left"> <img src="docs/figures/clrernet.jpg" height="200"\></p> <p align="left"> <img src="docs/figures/laneiou.jpg" height="160"\></p>Performance
CLRerNet achieves the <b>state-of-the-art performance on CULane benchmark </b> significantly surpassing the baseline.
| Model | Backbone | F1 score | GFLOPs |
|---|---|---|---|
| CLRNet | DLA34 | 80.47 | 18.4 |
| CLRerNet | DLA34 | 81.12±0.04 <sup>*</sup> | 18.4 |
| CLRerNet⋆ | DLA34 | 81.43±0.14 <sup>*</sup> | 18.4 |
* F1 score stats of five models reported in our paper. The release models' scores are 81.11 (CLRerNet) and 81.55 (CLRerNet⋆, EMA model) respectively.
Install
Docker environment is recommended for installation:
docker-compose build --build-arg UID="`id -u`" dev
docker-compose run --rm dev
See Installation Tips for more details.
Inference
Run the following command to detect the lanes from the image and visualize them:
python demo/image_demo.py demo/demo.jpg configs/clrernet/culane/clrernet_culane_dla34_ema.py clrernet_culane_dla34_ema.pth --out-file=result.png
Test
Run the following command to evaluate the model on CULane dataset:
python tools/test.py configs/clrernet/culane/clrernet_culane_dla34_ema.py clrernet_culane_dla34_ema.pth
For dataset preparation, please refer to Dataset Preparation.
Frame Difference Calculation
Filtering out redundant frames during training helps the model avoid overfitting to them. We provide a simple calculator that outputs an npz file containing frame difference values.
python tools/calculate_frame_diff.py [culane_root_path]
Also you can find the npz file [here].
Train
Make sure that the frame difference npz file is prepared as dataset/culane/list/train_diffs.npz.<br>
Run the following command to train a model on CULane dataset:
python tools/train.py configs/clrernet/culane/clrernet_culane_dla34.py
Speed Test
Calculate fps by inference iteration.
python tools/speed_test.py configs/clrernet/culane/clrernet_culane_dla34.py clrernet_culane_dla34.pth --filename demo/demo.jpg --n_iter_warmup 1000 --n_iter_test 10000
Citation
@article{honda2023clrernet,
title={CLRerNet: Improving Confidence of Lane Detection with LaneIoU},
author={Hiroto Honda and Yusuke Uchida},
journal={arXiv preprint arXiv:2305.08366},
year={2023},
}