confusion about the reweight for CD loss

[drilistbox]

Thanks for your greate work. And when I study the code, I find a re-weight, i.e. gt_pts_weights, is used for loss_pts between gt_pts and gt_paired_pts. But there is no re-weight for loss_pts between pred_pts and pred_paired_pts. Does it affect the performance?

    # calculate loss pts
    loss_pts = pred_pts.new_tensor(0)
    loss_pts += self.loss_pts(gt_pts,
                              gt_paired_pts,
                              weight=gt_pts_weights[..., None],
                              avg_factor=gt_pts.shape[0])
    loss_pts += self.loss_pts(pred_pts, 
                              pred_paired_pts,
                              avg_factor=pred_pts.shape[0])

[jbwang1997]

We did some experiments and observed that adding weight on loss_pts between pred_pts and pred_paired_pts has less effect on final performance.

[drilistbox]

第一层预测的点很少，只有600个，影响不大。但是到后面的decoder层，比如第6层，预测点数变为600x128=94800，不可能所以预测点都落到有物体的体素内，距离误差应该有不少在0.2外的吧。如果这会都不需要算下距离权重，说明权重对loss_pts(between pred_pts and pred_paired_pts)的影响不大。

但我有点迷惑，为啥权重对loss_pts(between gt_pts and gt_paired_pts)有影响呢，可以给一些启发么

[jbwang1997]

我认为可能还是 camera mask 的问题。pred_pts 因为 pred_pts 有很多还是在 camera mask 以外，因为这些区域看不见所以其实这些预测也并不算错，直接给所有 pred_pts 增加权重容易让模型梯度很大。我们还尝试只给在 camera mask 以内的 pred_pts 增加权重，但效果一般最后在放出来的版本里没有加。

而对于 gt_pts 与 gt_paired_pts 之间 loss 的加权，其实相当于直接让空的 gt_pts 拉过来一个 gt_paired_pts，这样很直接能增加 mIoU。