Dataset Used for Training the ICON Model

[caiocvsilva]

Hi,

I am trying to understand the datasets used for training the ICON model, and I'm a bit confused by what I've found in the paper compared to the provided code/documentation.

In the ICON paper, it is mentioned that the model was trained using 450 Renderpeople scans, which are a subset of the AGORA dataset. Additionally, the paper states that the THuman dataset was only used to evaluate the model's performance.

However, when looking at the code, it seems that the THuman2.0 dataset is used for training. The dataset.md document also mentions downloading and preparing the THuman2.0 dataset for training purposes, but there is no mention of how the AGORA dataset is used, nor are there instructions for preparing or using it.

Could you please clarify the following:

1. Was the ICON model trained using a subset of the AGORA dataset (specifically the 450 Renderpeople scans) as mentioned in the paper, or was THuman2.0 used for training instead?
2. If AGORA was indeed used, could you provide guidance or instructions on how to prepare and use the AGORA dataset for training the ICON model?
3. Is there any specific preprocessing or subset selection needed for the AGORA dataset that I should be aware of?

I would greatly appreciate any clarification on these points to better understand the training setup.

Thank you for your help!