Objects and Road Elements
What kind of labels will you provide for objects and road elements ?
At this stage we're only planning to provide online perception autolabels (as opposed to offline-optimized or human-in-the-loop ground truth labels), though we are investigating avenues to release higher quality labels especially to support, e.g., radar perception applications for which comparatively fewer datasets exist.
Reopening this issue for discoverability during the initial release period for this dataset.
Hi, could you provide some more detail about the auotlabels. The images have blurred_boxes with fields (frame_index, x1, y1, x2, y2) are there any labels? The lidar doesn't seem to have any autolabels, so I am assuming there are no 3D labels?
The only type of autolabels currently published is ego vehicle trajectories (under labels/egomotion) -- additional object and road labels are coming soon. I should note that these will be 3D labels in the local frame and may require projection, e.g., into each camera's image plane using the provided sensor extrinsics + intrinsics. We aim to provide a demo of that in the upcoming developer kit. Admittedly there's a lot of "coming soon" that you'll have to take my word on: depending on the application some folks (e.g., those interested in pure E2E imitation learning) can already get started, but for others it may be best to check back in a week.
The blurred boxes are provided for researchers that would like the exact details on how the original sensor videos were anonymized (i.e., modified to remove PII including faces and license plates -- you may have noticed this if you've opened up one of the videos for inspection), in case they'd like to use this information to mitigate any associated bias. I wouldn't consider these to be "labels" from an AV ML applications perspective; these boxes are provided for completeness but I expect that most folks will ignore them.