Questions about encoding

[smandava98]

Hi. I have some questions on how you encoded the SMPL parameters.

• If you already have it in 229 dim vector form, what is the case for reducing the dimensionality to 128?
• How was the SMPL predictions from PHALP encoded into a 229 dim vector in the first place in the dataset? I assume it was simple concatenation but wanted to know if any additional processing was done.
• Why subtract the SMPL poses by mean and standard deviation instead of directly using the PHALP predictions? Wouldn't that cause a skew in the image space?

[brjathu]

If you already have it in 229 dim vector form, what is the case for reducing the dimensionality to 128? --- just to have control over model size, this way we can train models with 128, 256, or other large-scale models.

How were the SMPL predictions from PHALP encoded into a 229 dim vector in the first place in the dataset? I assume it was a simple concatenation but wanted to know if any additional processing was done. --- yes, just simple concatenation. 207 (23 joints * 3*3) + 9 for global orientation + 10 for shape.

Why subtract the SMPL poses by mean and standard deviation instead of directly using the PHALP predictions? Wouldn't that cause a skew in the image space? --- It helped train the pose model with zero-centered poses, probably because the weights are initialized to support zero-centered inputs.

Wouldn't that cause a skew in the image space? --- I don't quite understand this part of the question.

[smandava98]

Thank you. By skew I mean you're shifting the X,y coordinates on the image right? That would impact the relationship between people in the camera frame