3D Bounding box quality in camera view is poor.

[al093]

Hi, I visualised couple of scenes from the dataset, and it looks like the position of 3D bounding boxes (BB) do not match the cars in image view quite often.

Some frames from the dataset:

• The left front car here:

• The right front car here:

• The left car here:

In some frames the boxes do match the cars but then it slowly accumulates a negative offset along the Y axis, or "lag behinds" the car in cam view. This quirk makes the dataset unusable for camera only algorithms.

My question here would be:

1. Is this expected ? I did verify that I am using the poses correctly, otherwise the boxes wont be accurate even occasionally.
2. Is this behaviour because the BB were labelled in lidar PC sparsely on some frames and then interpolated ? This might explain why the box positions seem to be correct in some frames but not in many.
3. Were the labelling done for every couple of frames, if yes, and if only those frames look better i would like to know a way to use only those for training and evaluation.

[al093]

Tagging @pfmark who I think might have clue. Any suggestions and comments would be helpful and appreciated :)

[mdfaheem786]

hi @al093, is it possible, to share your code of image visualization... as my viz is too bad, - https://github.com/scaleapi/pandaset-devkit/issues/148 thank you in advance,

[al093]

@mdfaheem786 Unfortunately i am not allowed to share it.

[pfmark]

@al093 can you share the scenes and time stamps you're looking at here?

[al093]

@pfmark Unfortunately I had to move on a long time ago because lack of community support and I did not note down the scene and timestamps. But anyways, this issue is visible in many many scenes so if you want to debug it, it wont be hard to find a scene with this issue. If you still need the scene for comparing visualizations let me know i will try to scavage what i had.