search

GuyTevet / MotionCLIP

Official Pytorch implementation of the paper "MotionCLIP: Exposing Human Motion Generation to CLIP Space"
MIT License
402 stars 38 forks source link

Why input size is 25 x 6? #19

Open kjw9899 opened 1 year ago

kjw9899 commented 1 year ago

what a great work!

I'm implementing with your code, but I have a question.

In your paper, Input is orientations in 6D representation of SMPL body model(24 x 6). When I check from debugging, The input shape is 25 x 6.

Which one is right? or What component is added to 24 x 6?

kjw9899 commented 1 year ago

I check that it seems to zero padding.

Can I ask why do this?

GuyTevet commented 1 year ago

The last channel is for global translation. Since its dimensions are 3 and not 6 it's zero-padded.

zhipenggong commented 1 year ago

where is global translation from, it seems to be from the first action2motion joint, but why action2motion switch 0# and 8#?

GuyTevet commented 1 year ago

If I remember correctly, this switching shouldn't be in use anymore. root rotation is at [0] and root translation is at [-1] (only the first 3 entries of it)

zhipenggong commented 1 year ago

The switching seems to be in use: https://github.com/GuyTevet/MotionCLIP/blob/main/src/datasets/amass.py#L94

GuyTevet commented 1 year ago

That's a definition, not a usage, I don't see where this one is being used.