Question about actual_depth

[menggui1993]

Hi, I have a question about the actual_depth processing before localnet. In function center2dtopc, the actual_depth is calculated for each grasp group. The value is rescaled by (1280 // output_size[0]). The depth is read from the depth image of size (1280, 720) directly, then added by the predicted depth_offset from anchornet. I can't see why another scaling factor (1280 // outputsize[0]) is needed. I referred to this part of code.
Could you please explain why? Thanks.

[ChenThree]

This is because we directly use the function pixel_depth_2_points for width conversion, which uses the original camera intrinsics with resolution (1280, 720). Thus, rescaling the depth, that is the z in the width conversion function, will result in the correct results. (But this may not be an ideal way for understanding)