Where is the code of adaptive face splitting?

[Zhangsihao-Yang]

I did not see the code that does the adaptive face splitting in GEOMetric.py. Could you guide me to the position?

[Zhangsihao-Yang]

I could see that part of the code in the old_GEOMetrics folder. But should it be used in the new folder?

[EdwardSmith1884]

No, I leave it out in the interest of large speed ups from vectorization.

[KnightOfTheMoonlight]

Could you make it public? @EdwardSmith1884. It looks like you treat the face split part as an end-to-end learning part for your whole project. However, after the face split, the adjacency matrix should be updated. If it is a post-process part, then It has no effect at all, cause this part will have not affect to the mesh deformation part.

[EdwardSmith1884]

@KnightOfTheMoonlight the code is public already here. Face splitting is not a post-processing step. It happens during learning. Quite a bit of work was put into making the adjacency matrix update somewhat efficient.

[KnightOfTheMoonlight]

@EdwardSmith1884 as your updated code does not contain the face slit part, your updated results from this comment https://github.com/EdwardSmith1884/GEOMetrics/issues/1#issuecomment-593428137 are obtained from no face split part too?

[EdwardSmith1884]

Yes this is with no space splitting.

[KnightOfTheMoonlight]

@EdwardSmith1884 Then this is against your ablation study part in your paper. Table 2. You have proven face split has improved the performance from 50.33 to 56.61 for the chair category.

[EdwardSmith1884]

No, this improvement is from:

• using different regularization parameters
• training with all classes simultaneously
• the increased training time afforded by increased speed of the model.

The face splitting was removed so that I could vectorize the training process, and make it as fast as possible.