search

akashsengupta1997 / STRAPS-3DHumanShapePose

Code repository for the paper: Synthetic Training for Accurate 3D Human Pose and Shape Estimation in the Wild (BMVC 2020)
MIT License
165 stars 18 forks source link

AMASS for training #9

Closed babyfisher closed 3 years ago

babyfisher commented 3 years ago

Hi, Thank you for sharing your great work! I noticed that you provided preprocessed AMASS to replace H3.6M for training. Did you use a subset of AMASS to get a similar data scale as H3.6M? What datasets in AMASS did you use? Do you use whole sequences or do you skip frames?

akashsengupta1997 commented 3 years ago

Hey sorry for the very late reply!

I believe I used the entirety of AMASS (or the version of it that was current at the time this repo was released). This did result in ~300k training poses, compared to around ~60k for H3.6M. However, I didn't seem to get much improvement in pose evaluation metrics despite using the larger training dataset, which is kind of curious. I haven't investigated why, but it could be related diversity of poses in AMASS vs H3.6M?

I sampled frames such that I obtained 2 training poses per second of motion capture.

babyfisher commented 3 years ago

Thank you for your detailed reply. I appreciate it.