search

geopavlakos / hamer

HaMeR: Reconstructing Hands in 3D with Transformers
https://geopavlakos.github.io/hamer/
MIT License
326 stars 28 forks source link

mano mean parameter #41

Closed luoxue-star closed 3 months ago

luoxue-star commented 4 months ago

May I ask how the mano_mean_params.npz file was obtained? Is it done by calculating the average of all poses and shapes? If this file is not present, what should the cfg.MODEL.MANO_HEAD.IEF_ITERS in the MANO HEAD forward be set to? Looking forward to your reply soon. Thanks!

geopavlakos commented 3 months ago

For the pose parameters, we use the mean hand pose from MANO (hands_mean). For the shape parameters, we inherit the average shape used by HMR networks (which is indeed computed as average across all body shapes from MoCap data).

The pretrained network expects this file to be available. If you don't want to use this file and you train a new model, I expect you will still get good performance regardless (maybe with a small drop).