Inconsistent Number of Motions Between This Pipeline and PHC When Processing AMASS Dataset

[Colin-Jing]

Hi! I am currently reproducing the first stage (Full Body Motion Tracking) and aiming to achieve the results reported in Table 1 of your paper. However, I noticed that the number of motions obtained after processing the AMASS dataset using your provided data processing pipeline is significantly less than the number obtained using the PHC data processing pipeline.
Could you please let me know if I might have made any mistakes during the processing?

[tesslerc]

The provided YAML files are based on the HumanML3D splits of the AMASS dataset. MaskedMimic uses this for learning text-to-motion.
We can upload the YAML files that patch the PHC structure.

Before we upload those, let's make sure they indeed span the same files. The train/test/validation split consists of 11316/137/346 motions each. Is that also what you see from PHC?

[Colin-Jing]

Thank you for your reply. I am looking forward to the upload of the YAML files for the PHC structure. That being said, I am a bit concerned that the current pipeline might not be able to properly process these YAML files, as HumanML3D might only be a subset of the PHC dataset. From my observation, it seems that the existing data processing pipeline cannot handle samples that are not present in HumanML3D.

Regarding the number of training and testing samples in PHC, as mentioned in the paper, PHC filters out low-quality samples from the AMASS dataset. According to the paper, the final numbers are 11,313 for the training set and 138 for the testing set.

When I run the data processing pipeline provided in PHC, I also get 138 samples for the testing set, which matches the paper. However, the total number of samples in the training and validation sets combined is 11,159, which is different from the number reported in the paper.

At the moment, I am not sure where the discrepancy comes from. You may also refer to these related issue:
ZhengyiLuo/PHC#5 (comment)
ZhengyiLuo/PHC#88 (comment)

[tesslerc]

There's no issue with loading motions that are not part of the HML3D dataset.
We'll upload the yaml files soon and update.

My bad on the numbers, they are zero-indexed. So test is also 138 on my end. Need to check possible discrepancies with the PHC pipeline and we'll upload.

[Colin-Jing]

Thank you so much! I really appreciate your help and look forward to your updates.