3D bbox projection to 2D

[mathild7]

Hi,

I am seeing that (as expected) the projected 3D bounding boxes on the image do not fit tightly.

This may be due to how the transformed 8pt bbox is downsampled into a 4pt bbox via this mechanism

# Compute the 2D bounding box and draw a rectangle x1 = np.amin(points[...,0]) x2 = np.amax(points[...,0]) y1 = np.amin(points[...,1]) y2 = np.amax(points[...,1])

Is there a better way to select these corners to get a tighter fitting bbox?

(Red -> Projected labels) (Blue -> Native labels)

[qingzi02010]

I have the similar question. As I learned, other public datasets only have 3D bounding boxes annotation, otherwise WAYMO dataset has both 3D and 2D bounding boxes. The 2D bounding boxes fit tightly, but the 3D bounding boxes can't. So I guess that 2D and 3D bounding boxes are annotated separately. Is there any suggestion to combine 2D bounding boxes and 3D bounding boxes, to make 3D bounding boxes fit better?

[DavidAdamczyk]

@peisun1115 Please can you explain how to find connection between 2D bounding box and 3D bounding box?

[DavidAdamczyk]

Maybe this is the solution: https://github.com/waymo-research/waymo-open-dataset/issues/414