To ensure the traceability, reproducibility and standardization for all ML datasets and models generated and consumed within TRI, we developed the Dataset-Governance-Policy (DGP) that codifies the schema and maintenance of all TRI's Autonomous Vehicle (AV) datasets.

• Schema: Protobuf-based schemas for raw data, annotations and dataset management.
• DataLoaders: Universal PyTorch DatasetClass to load all DGP-compliant datasets.
• Visualizer: Simple web-based visualizer for viewing annotations.
• CLI: Main CLI for handling DGP datasets.

Getting started is as simple as initializing a dataset-class with the relevant dataset JSON, raw data sensor names, annotation types, and split information. Below, we show a few examples of initializing a Pytorch dataset for multi-modal learning from 2D bounding boxes, and 3D bounding boxes.

from dgp.datasets import SynchronizedSceneDataset

# Load synchronized pairs of camera and lidar frames, with 2d and 3d
# bounding box annotations.
dataset = SynchronizedSceneDataset('<dataset_name>_v0.0.json',
    datum_names=('camera_01', 'lidar'),
    requested_annotations=('bounding_box_2d', 'bounding_box_3d'),
    split='train')

A list of starter scripts are provided in the examples directory.

• examples/load_dataset.py: Simple example script to load a multi-modal dataset based on the Getting Started section above.

You can build the base docker image and run the tests within docker via:

make docker-build
make docker-run-tests

Run streamlit-based interactive visualizer via:

make docker-start-visualizer