[fix] - bugs in shape decoration

[saadejazz]

• normalization should be along channel dimensions
• bg_mask is the inverse of inp_mask

[saadejazz]

Hi,

Thank you for open-sourcing your work. While reproducing the results, I came across a few potential bugs that you may want to look into.

I also noticed that the method does not seem to achieve strong visual performance except when using the precomputed demo features. In particular, when I recompute the demo features using the provided code, the same demo no longer aligns properly, despite having a clear and unoccluded view.

Could you please provide some context on this issue? Could be that the adaptation of Diff3F is not consistent?

Thank you.