problem about loss

[kellenf]

Thank you for this nice work~. I have some questions?

1. why to scale y/z ? but without x? how about this benefit?
2. some question around this distance formulation.
   inter_diag / (c_diag).
   Distance-IoULoss, pointed this formulation will caused large gradient.
   And I think replacing inter_diag by smooth_l1_loss about (cx, cy, cz, c_orientation) will be better?

It's grateful to receive your answer~

[hlsheng1]

Thank you.

1. I did a lot of experiments on kitti and found that this setting is better. These parameters can be adjusted, and the performance difference is not large.
2. Have not considered yet, I will do some experiments about this and response u later.

[kellenf]

Thanks for you responds! Now I have a another question.

1. In rdiou formulation, the delta_theta is [-1, 1]. but the delta_(cx, cy, cz) is not in [-1, 1]. Does this is a deviation in value-scale? I mean maybe we need a better encode way, how do you think about this?
2. I visualization the fig(x_axis->scores, y_xais->ious), but it's not a distribution around y = 1 * x as the QFL want it to be. I use accurate 3D IoU in my process. Did you do some exps about this topic?

[hlsheng1]

Thanks for your interest.

1. U mean making delta_theta and delta_(cx, cy, cz) in a same scale? I don't undersatnd the insight here. Why we must do this?
2. RDIoU should be better than 3D IoU, and 3D IoU should be better than the non-IoU optimization. Could you provide more information? U mean the visualization of using 3D IoU is not so good?