metric_scale

Open BIn-999 opened this issue 3 months ago • 1 comments

I found that the metric_scale prediction result has no corresponding supervision function. Doesn’t this lead to scale errors in the estimation?

Oct 27 '25 03:10 BIn-999

The affine invariant loss calculates scale, but the inference scale branch is unsupervised. I'm curious how the MLP branch of the model ensures scale consistency. Furthermore, the affine invariant loss is calculated as pred_points * scale + shift, while the inference branch is calculated as (pred_points + shift) * metric_scale. I'm confused as to how this corresponds.

Oct 27 '25 03:10 BIn-999