Oct13thNote.txt

Inside of EmocaProcessed_38k/geometry_detail and EmocaProcessed_38k/inputs:
- We have 38k cutoff of the original 41k Affectnet Images that are balanced by classes and results.
The cutoffs have the Flame render in "geometry_detail" and the simple processed original images cut around the faces in
"inputs"

- As of 13/10/2024, we want to modify this to have a Render on a image canvas sized like the original as if it could be
superimposed on top, like it does with the inputs.
- We can then process the original, for depth, semantics, and skeleton images, that would be used on a finetune.
- We also need to modify the loss to incorporate the EmonetLoss.

 
- The Semantic mapping is not really useful for faces as of right now.
We will use the following depth mapping: https://github.com/isl-org/ZoeDepth
We will use the following for face alignment: https://github.com/1adrianb/face-alignment
This is already used by EMOCA so it might be a repeat?