Currently we have explored various UV maps, and found out some particular artifacts that we believe is closely related to your choice of UV texture mappings. For example, we observed frequent color inconsistency when testing with radvani UV maps as showcased here (note the colors of both hands on bottom-right corner):
We believe the segmentation of body parts and the placement of UV atlases plays an important role in training, but so far we haven't reached a solid conclusion yet. Therefore we encourage the users to explore and report results trained with their own UV maps. This document hereby details the necessary steps to use your customized UV maps.
To create a UV map, you need to load an SMPL template model into any 3D animation software like MAYA or Blender. Then follow necessary steps to cut the mesh into different parts and unfold each part onto an UV grid map. Then assign 3D vertex to UV vertex correspondences and UV texture coordinates between [0,1].
To use your UV map for our projects, save your UV data as a single .obj file, name it {your_uv}_template.obj. Put it under data_utils folder and follow the instructions in PREPS.md to prepare datasets and labels.
But before that, it's highly recommended to check if your UV info is correct. You can run data_utils\uv_map_generator_unit_test.py for a quick check, it reads in first 10 items of your dataset (I'm using human36m) and create corresponding UV labels inside _test_cache folder.
The correct UV maps should look like this:
While an incorrect UV map could look like this (notice the color inconsistency inside each atlas):
