Some questions about training about icon HOT 7 CLOSED

These doubts are Crucial issues when I train and test ICON. Thanks for your help again, which will better help me to learn and reproduce ICON.

from icon.

Thanks for your reminder.
I have implemented retraining on the DeepHuman dataset and THuman2.0 dataset using the SMPL model and semantic map.
It finds that the features of ICON are extremely efficient, especially SMPL features, so that excellent models can be reconstructed with so few features. Your work has also inspired me how to extract pixel features and spatial features efficiently.
Thank you again.

from icon.

GT SMPL is used for implicit network training, which is the same case for normal network training.

Actually, the normal and implicit networks can be trained separately yet simultaneously, I used GT rendered normal images to train the normal network instead of the predicted normal images.

from icon.

Thanks for your reply. I maybe know some details.
"I used GT rendered normal images to train the normal network instead of the predicted normal images."---I don't quite understand this sentence.
GT rendered normal images are surely used for training the normal network since they serve as labels to supervise training.
However, If the normal and implicit networks are trained separately, which image is chosen during the training of the implicit networks, GT rendered normal images or predicted normal images from the normal network?

In addition, I would like to ask a few questions about testing and evaluation.
I have tested in THuman2.0 with your official version and these results reconstructed from ICON are very great in qualitative evaluation. For example, fingers and face details are even obtained. However, their quantitative metrics（including CD PSD l2 COS）are terrible. I would like to know if you encounter such a case that qualitative results outperform but quantitative metrics are terrible?

from icon.

It is GT rendered normal images which are used to train the implicit networks.

I didn't train ICON with THuman2.0 actually, I used AGORA. But according to your testing results, I thought you should first try to use GT-SMPL as input and see the final results.

Notably, the numbers in our paper, are conditioned on GT-SMPL or GT-SMPL with perturbed noise.

from icon.

Now both dataset preprocess code and training code are ready!

Dataset processing
Training

from icon.

@sunjc0306 New cloth-refinement module is released. Use -loop_cloth 200 to refine ICON's reconstruction, making it as good as the predicted clothing normal image.

from icon.