Release Data Processing Code or Detailed Annotations for TAPVid-3D Dataset

[Luo-Yihang]

Hi Skanda,

Thank you for the wonderful work on TAPVid-3D — it's been incredibly inspiring.

I was wondering if it would be possible to release the data processing code, or alternatively, provide more detailed annotations for the datasets (e.g., depth, dense tracks), so that the dataset can be more useful for other downstream tasks.

I have came across this related issue: #126, which includes some code snippets related to DriveTrack. Still, having a more comprehensive version or official release would be very helpful.

[skoppula]

Hi Luo!

Sorry for the late response! We didn't have plans to release much of the data processing code to be honest, as it takes a bit of time to get things reviewed for release. What annotations specifically would you want for your purposes, and on what subset split? Note that for some splits, e.g. Waymo, it's not possible to get dense tracks for the entire frame.

[Luo-Yihang]

Hi Koppula! Would it be possible to release the depth prediction for all subsets? I think it could be very valuable for the 3D tracking task. Thank you!

[skoppula]

I don't think we'd be able to get dense depth for all subsets -- e.g. the LIDAR in the Waymo subset provides sparse point clouds. Would providing that help you out?

[Luo-Yihang]

While sparse point clouds might not be necessary, would it be possible to release the depth for the other two datasets? Thank you!