llorz / SGA19_zoomOut

code for the paper "ZoomOut: Spectral Upsampling for Efficient Shape Correspondence"
https://arxiv.org/abs/1904.07865
44 stars 9 forks source link

about code #7

Open njnj0314 opened 2 years ago

njnj0314 commented 2 years ago

Hello, I am using your code to optimize my stationary point mapping matrix, but the following situation occurs, the first picture is the visualization before using zoomout, the second picture is after using, the color at the back is mapped to the front, a total of Of the 190 pairs of test results, about 30 of them have such a situation, why? (1): image (2): image

llorz commented 2 years ago

Hi,

The reason you get a back-to-front map after running ZoomOut might be that you start at a small functional map size. For example, if you start at dimension 2-4, you might get a left-to-right map, and you start ; if you start before 20, you might get a back-to-front map. (This is my experience for human shapes, you might get different results coming from symmetry ambiguity for other shapes). The reason is that the (global) shape symmetry is encoded in the first a few eigenfunctions with relatively small eigenvalues. You can find more discussions in this paper: https://arxiv.org/abs/2006.02532

As an easy fix, you can try ZoomOut starting from 20 for human shapes, I think this should be resolved.

Hope this helps, Best regards, Jing