search

argoverse / argoverse-api

Official GitHub repository for Argoverse dataset
https://www.argoverse.org
Other
833 stars 239 forks source link

Object Label not Continuous Across Frames #278

Open ziqipang opened 2 years ago

ziqipang commented 2 years ago

Hi,

Recently I am working on the tracking dataset, and found that the labeling of objects is not continuous across frames. Specifically, an object A may be labeled on frame 2 and 4, but missed on frame 3. I am wondering if my observation is wrong, and if so, I am interested in what leads to this design?

Thank you so much!

johnwlambert commented 2 years ago

Hi @ziqipang, thanks for your interest in our work.

Is there a specific sequence (with frame IDs and object IDs) you could refer us to, so we can take a look at the log you are referring to?

This may be due to occlusion. For example, if a pedestrian is temporarily completely occluded (from the AV's perspective) by another vehicle, it is not possible to label that pedestrian in 3d, as we would have no observations about the pedestrian's location for some period of time (although there is, of course, a prior that their movement would be stationary or linear).

ziqipang commented 2 years ago

Thank you for the reply! I have some incomplete list of IDs on the sample tracking sequence. Please check on the following objects: 6c5faf3e-04ed-4f45-94e9-bb57b345f6eb 2af61a36-20b0-4bfe-b341-dc4d03548ead f0ccac2d-4dd6-44a8-9d18-feb030186149 371947d1-c28c-415a-924b-e39500b4f961 e7297ebe-a5af-4be3-ae61-bc28ec3fcdfa 26731792-9dc0-483f-8b08-63e50b19f85a

Thank you so much!

johnwlambert commented 2 years ago

Thanks. Are these from one particular log, or from multiple?

ziqipang commented 2 years ago

They are all from the sample point cloud sequence, and the IDs above are the track_id, belonging to different objects (tracklets). Hopefully, this time my expression is rigorous lol.