Waymo-Toolkit

This is a toolkit for extracting elements (e.g. images, lasers and labels, which modified from offical dataset.proto defination to annotation.proto) from Waymo Open Dataset and visualizing the extracted images, laser points and labels with OpenCV and Mayavi

Installation

git clone https://github.com/DapengFeng/waymo-toolkit.git
cd waymo-toolkit
conda create -n waymo-toolkit python=3.7 vtk mayavi tabulate matplotlib numba scipy
conda activate waymo-toolkit
python setup.py install

Extraction

/path/to/waymo_dataset
|---training
|   |---*.tfrecord
|---validation
|   |---*.tfrecord
|---testing
|   |---*.tfrecord

1. Extract all images

python tools/extract.py --source=/path/to/waymo_dataset --dest=/path/to/save_dir --type=all --image

2. Extract all laser points

python tools/extract.py --source=/path/to/waymo_dataset --dest=/path/to/save_dir --type=all --laser

3. Extract all labels and randomlly select 10% frames as the subset from the whole dataset. And I use the default random seed is 20200319, which is release date of waymo open dataset v1.2. You can change it to your own one, using the flag--seed

python tools/extract.py --source=/path/to/waymo_dataset --dest=/path/to/save_dir --type=all --label --subset

After the above steps, you can see the script will generate the following folders in the /path/to/save_dir. If you only want to extract the training data, you can change the flag --type=all to --type=train.

/path/to/save_dir
|---training
|   |---image_0 (FRONT)
|   |---image_1 (FRONT_LEFT)
|   |---image_2 (FRONT_RIGHT)
|   |---image_3 (SIDE_LEFT)
|   |---image_4 (SIDE_RIGHT)
|   |---laser (x,y,z)
|   |---laser_r (range)
|   |---laser_i (intensity)
|   |---laser_e (elongation)
|   |---label
|   |---split.txt
|---validation
|   |--- the same as training
|---testing
|   |--- the same as training (except the split.txt)

Visualization

Visualize the image

python tools/visualize.py --source=/path/to/save_dir --image --project

python tools/visualize.py --source=/path/to/save_dir --image --box2d --project

python tools/visualize.py --source=/path/to/save_dir --image --box3d

Visualize the point cloud.

python tools/visualize.py --source=/path/to/save_dir --laser

License

Waymo-Toolkit is released under the Apache 2.0 license.

Citing Waymo-Toolkit

If you use Waymo-Toolkit in your research, please use the following BibTeX entry.

@misc{feng2020waymo-toolkit,
  author =       {Dapeng Feng},
  title =        {Waymo-Toolkit},
  howpublished = {\url{https://github.com/DapengFeng/waymo-toolkit}},
  year =         {2020}
}