Code for 2020 China Hualu Cup Data Lake Algorithm Competition (Lane Detection Track)

Introduction

This repo contains the first place solution by the Incredible@NTU team for the Lane Detection Track of the 2020 China Hualu Cup Data Lake Algorithm Competition. For more information about this task, we refer you to the official website. In this competion, We adopt the ERFNet as our base model for its lightweight and large receptive field, which are critical factors in this task. Furthermore, we perform data cleaning, sky removing and weighted crossentropy loss as the main techniques along with several other tricks to achieve the compelling performance (42.25 mIoU on Final testB) in a highly efficient manner. No model ensemble is used in our solution. Finally, our solution won the champion out of 576 teams in this competition. An overview of our solution is available here (passwd: shid).

Updates

• Add visualization.py for label visualization. See utils/README.md for details. (04/01/2021)

Contents

1. Installation
2. Datasets
3. Training
4. Evaluation

Installation

• Anaconda3
• PaddlePaddle 1.8.5
• OpenCV 4.2.0.32
• CUDA 10.0 or 10.1
• cuDNN 7.6.5
• TensorBoard (optional)
• pycocotools (optional)

Notes:

• It may also run well on other versions but with no guarantee.

• TensorBoard is optional, just for recording training process. If not installed, feel free to comment relevant lines of code.

• pycocotools is only required for visualization in utils/visualization.py.

For your convenience, we wrap up the installation process with the following commands.

conda create -n paddle python=3.7 -y && conda activate paddle
pip install opencv-python==4.2.0.32
python -m pip install paddlepaddle-gpu==1.8.5.post107 -i https://mirror.baidu.com/pypi/simple
pip install pycocotools

Datasets

We first construct our training set by combining the training data released in the preliminary and final round, which consists of 15,503 training images in total. After that, data cleaning and preprocessing techniques (as described below) are conducted to obtain higher quality training samples, resulting in 15,170 training images finally.

1. We manually filter out or re-label those wrongly annotated images. It can be roughly divided into three types, i.e., wrong annotation, wrong class label and miss label. For wrong annotation, it's shown on the left that a weird triangle (green) appears in the middle of the image with no actual lanes annotated. In the middle, the left-dashed-right-solid line (highlighted in yellow) is annotated as left-solid-right-dashed one, which is deemed as wrong class label. The last image belongs to miss label case since the dashed yellow line is missed.
2. There also exists some duplicate images with different annotation label maps. In this case, we manually select the one we regard owns the best annotation map and ignore all the rest.
3. We resize all the images to the same size of 1280*720 to reduce data loading time.

We assume the directory layout for the competition dataset PreliminaryData as below.

Project_ROOT
└── datasets
    ├── __init__.py
    ├── lane_det.py
    └── PreliminaryData
        ├── train_pic            # contains training images
        │   ├── 10008283.jpg
        │   ├── ...
        │   └── 10024760.jpg
        ├── train_label          # contains correponding grayscale labels
        │   ├── 10008283.png
        │   ├── ...
        │   └── 10024760.png
        └── testB                # contains test images
            ├── 10014001.jpg
            ├── ...
            └── 10016129.jpg

Note: Due to copyright issues, we do not release the dataset used for the competition, feel free to adopt our method on your own dataset.

Training

To train the model, run the following command.

python train_erfnet_paddle.py --epochs 150 -b 8 --lr 0.01

By default, we train our model with batch size 8 for 150 epochs on one RTX 2080ti GPU, which takes up approximately 7,000 MB GPU memory and 24hrs to finish. Our model starts with the pretrained weights pretrained/ERFNet_pretrained.pdparams, which is converted from the released torch version pretrained weights on Cityscape dataset.

Notes:

• If you wish to reproduce our results, please do not modify the batch size as it might impose uncertainty influence on the optimization steps.

• It will save the outputs (models and other records) to trained/ by default. Feel free to change it to your expected directory by specifying --save-dir path/to/output.

Evaluation

To reproduce our final results, just run this command:

python test_erfnet_paddle.py

By default, it will test on our released model trained/ERFNet_trained.pdparams, which exactly obtains the final results we submitted to the evaluation server.

To evaluate the models trained by yourself, simply add the --resume augment as:

python test_erfnet_paddle.py --resume trained/erfnet_epxxx

Replace xxx with the epoch point you want to test. Note that we recommend you to test the model with the highest mIoU on the training set, which is indicated by the file trained/best_model.

The testing results (prediction maps of lane markings) will be saved in results/result by default.

License

This repo is released under the Apache 2.0 License (refer to the LICENSE file for details).

Acknowledgement

This repo is mainly based on Codes-for-Lane-Detection, many thanks to them.

Should you have any questions regarding this repo, feel free to email me at ze001@e.ntu.edu.sg.