Training with MMPD

[znygithub]

Why are both DiffNormalized and Standardized during preprocessing when using MMPD as the training set?

[yahskapar]

Hi @znygithub,

This is something specific to the neural method you're using (e.g., TS-CAN) than the dataset - we typically adopted whatever the defaults in the provided configs are for fairer comparisons in our NeurIPS 2023 Datasets and Benchmarks paper.

[znygithub]

Okay, thank you for your answer. I have reviewed the data processing for each model.

I found the processing of EfficientPhys to be quite strange. In YAML SETTING, video is standardized, but the labels are diff-normalized. Wouldn't this cause misalignment between the video and labels? I checked the PhysNet model code, and it seems that the model itself does not perform diff-normalization internally.

Thanks for your answer

DATA_TYPE: ['Standardized']，LABEL_TYPE: DiffNormalized

[yahskapar]

@znygithub,

That all sounds expected - EfficientPhys has a normalization module internally (pictured below) that effectively avoids any misalignment that you're suggesting.