Closed gushengbo closed 1 year ago
The training instruction of normal network could be found at training.md
Could you share some bad results of your bad normal maps? Current pre-trained normal network checkpoint (normal.ckpt
) is trained on a larger internal dataset, but promising results should be achieved even only trained on THuman2.0
@YuliangXiu This is the result:
I can succesfully reconstruct it in front, but not the back.
sorry, I have another question. Could I train ICON or ECON with multiple gpus?
@YuliangXiu This is the result:
![]()
I can succesfully reconstruct it in front, but not the back.
This is normal, it's much harder to estimate back normal than front normal. This is because only the front RGB image is provided during the inference, so though the back normal is conditioned on SMPL-X back normal, it's almost hallucinated.
sorry, I have another question. Could I train ICON or ECON with multiple gpus?
Yes, you can check PyTorch-Lightning for more details.
o
Yes ,thank you. So just train back normal net with Thuman2 may get unperfect result.
o
Yes ,thank you. So just train back normal net with Thuman2 may get unperfect result.
Hi, I find that you train ECON and ICON with Thuman2 in Experiment. However, when I use thuman2 to train the back normal net in ICON, I got a bad result. Could you tell me how you train ICON with Thuman2? What are the training details?