WaymoCOCO

This converter converts the Waymo Open Dataset to COCO format. Current implementation supports to extract the 2D information of the dataset v1.2 (the version used in the Waymo Open Dataset challenge at CVPR 2020).

Installation

Requirements

• Linux
• Python 3.6+
• TensorFlow 1.15.0, 2.0.0, 2.1.0

Example using conda

conda create -n waymococo python=3.7
conda activate waymococo
pip install tensorflow==2.1.0
git clone https://github.com/shinya7y/WaymoCOCO.git
cd WaymoCOCO

Download

1. Access https://waymo.com/open/download/ .
2. Fill in the form and check the license for user registration.
3. Download the Training, Validation, and Test sets to ${HOME}/data/waymotfrecord/training/, ${HOME}/data/waymotfrecord/validation/, and ${HOME}/data/waymotfrecord/testing/, respectively.

You can download them quickly by the commands below. Installing Google Cloud SDK may be needed in advance.

gcloud auth login
# follow messages for authentication

gsutil -m cp -r gs://waymo_open_dataset_v_1_2_0_individual_files/training/ ${HOME}/data/waymotfrecord/
gsutil -m cp -r gs://waymo_open_dataset_v_1_2_0_individual_files/validation/ ${HOME}/data/waymotfrecord/
gsutil -m cp -r gs://waymo_open_dataset_v_1_2_0_individual_files/testing/ ${HOME}/data/waymotfrecord/

Conversion

WaymoCOCO f0 (frame 0)

The Waymo Open Dataset is large, but for many cases, it's too large. Using its subsets is useful when you would like to:

• do much trial and error before full training.
• evaluate the generalization of your method on the second dataset other than COCO.

This converter supports to extract 1/10 size dataset based on the ones place of frame index (e.g., frames 0, 10, 20, ..., 190).

# convert val
python convert_waymo_to_coco.py \
    --tfrecord_dir ${HOME}/data/waymotfrecord/validation/ \
    --work_dir ${HOME}/data/waymococo_f0/ \
    --image_dirname val2020 \
    --image_filename_prefix val \
    --label_filename instances_val2020.json \
    --add_waymo_info \
    --frame_index_ones_place 0
# convert train
python convert_waymo_to_coco.py \
    --tfrecord_dir ${HOME}/data/waymotfrecord/training/ \
    --work_dir ${HOME}/data/waymococo_f0/ \
    --image_dirname train2020 \
    --image_filename_prefix train \
    --label_filename instances_train2020.json \
    --add_waymo_info \
    --frame_index_ones_place 0
# convert test
python convert_waymo_to_coco.py \
    --tfrecord_dir ${HOME}/data/waymotfrecord/testing/ \
    --work_dir ${HOME}/data/waymococo_f0/ \
    --image_dirname test2020 \
    --image_filename_prefix test \
    --label_filename image_info_test2020.json \
    --add_waymo_info \
    --frame_index_ones_place 0

WaymoCOCO full

Full conversion is also available. Please note that a machine with 208-416 GB of CPU memory is needed for full training in the case of MMDetection v2.0.

# convert val
python convert_waymo_to_coco.py \
    --tfrecord_dir ${HOME}/data/waymotfrecord/validation/ \
    --work_dir ${HOME}/data/waymococo_full/ \
    --image_dirname val2020 \
    --image_filename_prefix val \
    --label_filename instances_val2020.json \
    --add_waymo_info
# convert train
python convert_waymo_to_coco.py \
    --tfrecord_dir ${HOME}/data/waymotfrecord/training/ \
    --work_dir ${HOME}/data/waymococo_full/ \
    --image_dirname train2020 \
    --image_filename_prefix train \
    --label_filename instances_train2020.json \
    --add_waymo_info
# convert test
python convert_waymo_to_coco.py \
    --tfrecord_dir ${HOME}/data/waymotfrecord/testing/ \
    --work_dir ${HOME}/data/waymococo_full/ \
    --image_dirname test2020 \
    --image_filename_prefix test \
    --label_filename image_info_test2020.json \
    --add_waymo_info

Other options

Please see convert_waymo_to_coco.py.

Creating symlinks (optional)

If you prepared WaymoCOCO f0 and WaymoCOCO full, the directory structure is as follows.

${HOME}/data
├── waymococo_f0
│   ├── annotations
│   ├── test2020
│   ├── train2020
│   └── val2020
└── waymococo_full
    ├── annotations
    ├── test2020
    ├── train2020
    └── val2020

Even if you use full training set for training, f0val is sufficient for validation. It is useful to create symlinks for that.

mkdir -p ${HOME}/data/waymococo/annotations
ln -s ${HOME}/data/waymococo_full/annotations/image_info_test2020.json ${HOME}/data/waymococo/annotations/image_info_test2020.json
ln -s ${HOME}/data/waymococo_full/annotations/instances_train2020.json ${HOME}/data/waymococo/annotations/instances_train2020.json
ln -s ${HOME}/data/waymococo_f0/annotations/instances_val2020.json ${HOME}/data/waymococo/annotations/instances_val2020.json
ln -s ${HOME}/data/waymococo_full/test2020 ${HOME}/data/waymococo/test2020
ln -s ${HOME}/data/waymococo_full/train2020 ${HOME}/data/waymococo/train2020
ln -s ${HOME}/data/waymococo_f0/val2020 ${HOME}/data/waymococo/val2020

If you use mmdetection, it is recommended to create symlinks in your mmdetection directory.

ln -s ${HOME}/data/waymococo_f0 ${MMDET_DIR}/data/waymococo_f0
ln -s ${HOME}/data/waymococo_full ${MMDET_DIR}/data/waymococo_full
ln -s ${HOME}/data/waymococo ${MMDET_DIR}/data/waymococo
# or simply
ln -s ${HOME}/data ${MMDET_DIR}/data

Acknowledgements

The files in waymo_open_dataset directory are borrowed from the official code to mitigate dependency. The official code and Waymo-Dataset-Tool (converter for KITTI format) were referred to write this converter.