search

shubham-goel / 4D-Humans

4DHumans: Reconstructing and Tracking Humans with Transformers
https://shubham-goel.github.io/4dhumans/
MIT License
1.24k stars 118 forks source link

Using SMIL instead of SMPL #44

Closed IotaCZE closed 1 year ago

IotaCZE commented 1 year ago

Hi! I was wondering whether it is possible to somehow easily exchange the SMPL model for SMIL to get better results on infants. I have just tried using the SMIL model (in place of SMPL), but it gives me completely the same results as with SMPL. I guess I am missing something... Thanks

geopavlakos commented 1 year ago

HMR2.0 predicts shape parameters for the SMPL model. The SMIL model has a different shape space, so you will not get something meaningful by just using the parameters from HMR2.0.

IotaCZE commented 1 year ago

So theoretically if I changed the output shape parameters to SMIL and retrained it on my dataset, I could get something meaningful?

geopavlakos commented 1 year ago

Yes, that could work, but you need to retrain or finetune on the appropriate data.