Extracting features from HumanML3D

[zhenglinpan]

• [x] Understand HumanML3D motion encoders
• [x] Understand FID calculation
• [ ] Extract intermediate representations

[zhenglinpan]

Thoughts: What's the point of extracting intermediate representations? -> Validate if Gaussian. Which layer? Unspecified. How to validate Multivariate Gaussian from 1-D vecs? Bias: with same structure, but different dataset. See FID and Gaussian. Gohar: which encoder is the best.?

[zhenglinpan]

email draft:

Hi Chuan:

Good afternoon Chuan :), I'm Aiden from Professor Li's class this term, I'm working on a project w.r.t FID metric, which is widely used in evaluating the result of human motion generation. This metric is also used in many of your works. I've read some of your them, they are absolutely impressive😎👍🏼

That said, I may have some other small question that I would like to have a discussion with you, it would be very kind of you to enlighten my queries.

Here are the facts I want to find out:

1. On which layer did you use FID in your work Action2Motion and Generating Diverse and Natural 3D Human Motions from Text? The original FID indicates How do you deal with sequence? How did you make the comparison? How many samples did you use?

[zhenglinpan]

Found the answers through reading codes & paper.

• The result was obtained from Snippet Code, as stated in paper
• If a motion clip contains 16 frames, FID was calculated on a batch of 32, 32*16=512 frames.
• Text-generated motion vs ground truth motion.
• A batch(32 in code).

[zhenglinpan]

• [x] Extract intermediate representations