Hi, I have a question about the Stage-2 synthetic annotations in navhard_two_stage.
When projecting the 3D boxes from synthetic_scene_pickles onto camera images with the official NAVSIM camera visualization path, many boxes appear vertically shifted or misaligned, while BEV-style visualization looks reasonable.
I checked the box fields and found that, in the sampled Stage-2 synthetic scenes, the z field appears to match the wrapped heading value. Since the documented box format seems to be:
[x, y, z, length, width, height, heading]
I wanted to ask whether the z field in Stage-2 synthetic pickles is intended to be valid for camera 3D bbox projection, or whether these annotations are only intended for BEV / planar evaluation.
For reference, in my sample:
- Stage-2 synthetic boxes:
2834/2834 had z == wrapped heading
- corresponding raw
navsim_logs/test boxes: 0/3708
- Stage-1 raw boxes from the same scene filter:
0/15891
Could you confirm the intended use of Stage-2 synthetic annotations for camera projection?
Hi, I have a question about the Stage-2 synthetic annotations in
navhard_two_stage.When projecting the 3D boxes from
synthetic_scene_picklesonto camera images with the official NAVSIM camera visualization path, many boxes appear vertically shifted or misaligned, while BEV-style visualization looks reasonable.I checked the box fields and found that, in the sampled Stage-2 synthetic scenes, the
zfield appears to match the wrappedheadingvalue. Since the documented box format seems to be:[x, y, z, length, width, height, heading]I wanted to ask whether the
zfield in Stage-2 synthetic pickles is intended to be valid for camera 3D bbox projection, or whether these annotations are only intended for BEV / planar evaluation.For reference, in my sample:
2834/2834hadz == wrapped headingnavsim_logs/testboxes:0/37080/15891Could you confirm the intended use of Stage-2 synthetic annotations for camera projection?