Question about depth completion

[canteen-man]

I follow the paper describes, find the depth completion algorithm in reference list named "In Defense of Classical Image Processing" and the code repo. And I try to complete the depth map by using that code. I found may be the 3D coordinate provide in the ONCE-3DLanes is not dense enough for that depth completion code. So I use the interplotation code in utils from ONCE3DLanes repo to interploate the 3D coordinate. Then project the 3D coordinate to 2D depth map with 16 bit value.(3D depth* 256 and round in uint16) The input depth map after interplote like this (set every points in 65535 for show, the truth input depth map is the truth depth value)

And I choose the "Fast" mode in that depth completion code. The output like this. I think the output depth map is not correct.

And when I modify the density of interpolation, the more dense depth map looks like below

The depth complete output is like this.

Do I need to modify the depth completion code or any step is wrong.

[sumenpuyuan]

hello I found that the kitti data set is also 3D depth* 256 and round in uint16, what is the significance of this data preprocessing

[sumenpuyuan]

Does the completion fail if blocked by a vehicle

[Ivanyan0516]

No depth completion when a lane point is blocked by a vehicle since we do not have the real z value.