Waymo Open Driving Data

Large fleet drives with high-resolution LiDAR and camera labels for 3D detection, tracking and motion forecasting.

Waymo Open pairs fleet LiDAR and camera frames with 3D tracks, so detection plus forecasting trains on long runs

Why Does This Exist?

Short clips test one frame well and motion poorly. Forecasting needs long, consistent runs with stable identities and maps. Waymo Open released fleet-scale drives with high-resolution LiDAR, cameras, 3D boxes, tracks and motion labels, so tracking plus forecasting could share one large ground.

This page covers sensor roles, track I/O and appropriate use..Verify the current release before quoting hours or scenes.

Think of It Like This

A taxi fleet with shared logbooks

One taxi log helps a little. A thousand taxis logging the same streets with the same forms lets you learn traffic habits, not just one ride.

Waymo Open is that fleet log. Long runs, same labels, room to study motion rather than snapshots.

How It Actually Works

Purpose is 3D detection, tracking and motion forecasting at fleet scale. Inputs are range images, LiDAR sweeps, camera frames and maps. Outputs are 3D boxes, consistent track IDs and future waypoints by task. Appropriate use is large-scale driving baselines and forecasting studies with heavy compute.

Toy track walkthrough

Take 10 truth boxes across frames with 1 miss, 1 false alarm and 0 switches. MOTA is $1 - (1 + 1 + 0) / 10 = 0.8$ . Identities stay clean, so IDF1-style association sits high while detection leaks, splitting the blame correctly.

Code

MOTA arithmetic that separates detection from association.

1def mota(miss: int, fp: int, switch: int, total: int) -> float:2    return round(1 - (miss + fp + switch) / max(total, 1), 3)3
4
5print(mota(1, 1, 0, 10))6# -> 0.8

One miss plus one false alarm on ten gives 0.8.

Watch Out For

Underestimating compute and storage

Fleet LiDAR plus long runs demand heavy IO and GPUs. The symptom is stalled training blamed on the model. Fix it by starting from sampled slices and range-image loaders before scaling to full runs.

The Quick Version

Waymo Open maps fleet frames to 3D boxes, tracks and forecasting labels.
Inputs are LiDAR plus cameras and maps, outputs are 3D tracks and waypoints.
Toy 1 miss and 1 false alarm on 10 gives MOTA 0.8.
Use it for large driving baselines and motion studies.
Verify release scale officially and budget compute early.