Skip to content
Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.

Codes for "Learning Lightweight Lane Detection CNNs by Self Attention Distillation"

This repo also contains Tensorflow implementation of "Spatial As Deep: Spatial CNN for Traffic Scene Understanding". (SCNN-Tensorflow)


  1. ERFNet-CULane-PyTorch has been released. (It can achieve 73.1 F1-measure in CULane testing set)

  2. ENet-Label-Torch, ENet-TuSimple-Torch and ENet-BDD100K-Torch have been released.

Key features:

(1) ENet-label is a light-weight lane detection model based on ENet and adopts self attention distillation (more details can be found in our paper).

(2) It has 20 × fewer parameters and runs 10 × faster compared to the state-of-the-art SCNN, and achieves 72.0 (F1-measure) on CULane testing set (better than SCNN which achieves 71.6). It also achieves 96.64% accuracy in TuSimple testing set (better than SCNN which achieves 96.53%) and 36.56% accuracy in BDD100K testing set (better than SCNN which achieves 35.79%).

(3) Applying ENet-SAD to LLAMAS dataset yields 0.635 mAP in the multi-class lane marker segmentation task, which is much better than the baseline algorithm which achieves 0.500 mAP. Details can be found in this repo.

(Do not hesitate to try our model!!!)

  1. Multi-GPU training has been supported. Just change BATCH_SIZE and GPU_NUM in, and then use CUDA_VISIBLE_DEVICES="0,1,2,3" python Thanks @ yujincheng08.



  1. Install necessary packages:
    conda create -n tensorflow_gpu pip python=3.5
    source activate tensorflow_gpu
    pip install --upgrade tensorflow-gpu==1.3.0
    pip3 install -r SCNN-Tensorflow/lane-detection-model/requirements.txt
  1. Download VGG-16:

Download the vgg.npy here and put it in SCNN-Tensorflow/lane-detection-model/data.

  1. Pre-trained model for testing:

Download the pre-trained model here.



The ground-truth labels of TuSimple testing set is now available at TuSimple. The annotated training (#frame = 3268) and validation labels (#frame = 358) can be found here, please use them (list-name.txt) to replace the train_gt.txt and val_gt.txt in Moreover, you need to resize the image to 256 x 512 instead of 288 x 800 in TuSimple. Remember to change the maximum index of rows and columns, and detailed explanations can be seen here. Please evaluate your pred.json using the labels and this script. Besides, to generate pred.json, you can refer to this issue.


The whole dataset is available at CULane.


The whole dataset is available at BDD100K.



cd SCNN-Tensorflow/lane-detection-model
CUDA_VISIBLE_DEVICES="0" python tools/ --weights_path path/to/model_weights_file --image_path path/to/image_name_list --save_dir to_be_saved_dir

Note that path/to/image_name_list should be like test_img.txt. Now, you get the probability maps from our model. To get the final performance, you need to follow SCNN to get curve lines from probability maps as well as calculate precision, recall and F1-measure.

Reminder: you should check and to ensure that the processing of image path is right. You are recommended to use the absolute path in your image path list. Besides, this code needs batch size used in training and testing to be consistent. To enable arbitrary batch size in the testing phase, please refer to this issue.


CUDA_VISIBLE_DEVICES="0" python tools/ --net vgg --dataset_dir path/to/CULane-dataset/

Note that path/to/CULane-dataset/ should contain files like train_gt.txt and val_gt.txt.


  1. TuSimple testing set:
Model Accuracy FP FN
SCNN-Torch 96.53% 0.0617 0.0180
SCNN-Tensorflow -- -- --
ENet-Label-Torch 96.64% 0.0602 0.0205

The pre-trained model for testing is here. (coming soon!) Note that in TuSimple, SCNN-Torch is based on ResNet-101 while SCNN-Tensorflow is based on VGG-16. In CULane and BDD100K, both SCNN-Torch and SCNN-Tensorflow are based on VGG-16.

  1. CULane testing set (F1-measure):
Category SCNN-Torch SCNN-Tensorflow ENet-Label-Torch ERFNet-CULane-PyTorch
Normal 90.6 90.2 90.7 91.5
Crowded 69.7 71.9 70.8 71.6
Night 66.1 64.6 65.9 67.1
No line 43.4 45.8 44.7 45.1
Shadow 66.9 73.8 70.6 71.3
Arrow 84.1 83.8 85.8 87.2
Dazzle light 58.5 59.5 64.4 66.0
Curve 64.4 63.4 65.4 66.3
Crossroad 1990 4137 2729 2199
Total 71.6 71.3 72.0 73.1
Runtime(ms) 133.5 -- 13.4 10.2
Parameter(M) 20.72 -- 0.98 2.49

The pre-trained model for testing is here. Note that you need to exchange the order of VGG-MEAN in and change the order of input images from RGB to BGR since the pre-trained model uses opencv to read images. You can further boost the performance by referring to this issue.

  1. BDD100K testing set:
Model Accuracy IoU
SCNN-Torch 35.79% 15.84
SCNN-Tensorflow -- --
ENet-Label-Torch 36.56% 16.02

The accuracy and IoU of lane pixels are computed. The pre-trained model for testing is here. (coming soon!)



If you use the codes, please cite the following publications:

  title={Learning Lightweight Lane Detection CNNs by Self Attention Distillation},
  author={Hou, Yuenan and Ma, Zheng and Liu, Chunxiao and Loy, Chen Change},
  journal={arXiv preprint arXiv:1908.00821},

  author = {Xingang Pan, Jianping Shi, Ping Luo, Xiaogang Wang, and Xiaoou Tang},  
  title = {Spatial As Deep: Spatial CNN for Traffic Scene Understanding},  
  booktitle = {AAAI Conference on Artificial Intelligence (AAAI)},  
  month = {February},  
  year = {2018}  

    title={Agnostic Lane Detection},
    author={Yuenan Hou},


This repo is built upon SCNN and LaneNet.


If you have any problems in reproducing the results, just raise an issue in this repo.

To-Do List

  • Test SCNN-Tensorflow in TuSimple and BDD100K
  • Provide detailed instructions to run SCNN-Tensorflow in TuSimple and BDD100K
  • Upload our light-weight model (ENet-SAD) and its training & testing scripts


Learning Lightweight Lane Detection CNNs by Self Attention Distillation (ICCV 2019)





No releases published


No packages published

Contributors 4