Question about DISCRETIZE in caddn.yaml

[zjyalice]

Hi~ thanks for sharing your great work!

I want to train CaDDN on custom dataset. I notice that the values of depth_min and depth_max in the DISRETIZE in caddn.yaml match the range of x in POINT_CLOUD_RANGE. However, the -y direction of the custom dataset in the lidar coordinate system is the forward direction. Therefore, 1) Do depth_min and depth_max refer to the distances in the advancing direction under the camera? Does it have to be numerically consistent with the range of the heading of POINT_CLOUD_RANGE? 2) How is the number of num_bins obtained? Is it set based on different datasets?

Thanks!

[codyreading]

Hi and thanks for the interest!

1) That's correct, they are distances in the advancing direction. The depth_min and depth_max set the max distances within the Frustum grid, while the parameters in POINT_CLOUD_RANGE set the range of the Voxel grid. The depths aren't required to be consistent, but it would be wasteful to set them to separate ranges, as this would lead to voxel grids with empty values (at depths not covered by the Frustum grid) or Frustum grid depths that aren't used in the voxel grid.

2) num_bins is set differently on different datasets. It is a purely empirical number, I would recommend a few settings and see which works best on the custom dataset. You might as well start with 80.

[zjyalice]

Thanks for your reply! By the way, have you evaluated the depth distribution predicted by the deeplabV3? How much is the gap between the result and depth map?

[codyreading]

I haven't done that evaluation, but qualitatively it looks quite accurate. I don't have specific metrics however.

[zjyalice]

Thanks for help!