search

dahiyaaneesh / peclr

This is the pretraining code for PeCLR. An equivariant contrastive learning framework for 3D hand pose estimation. The paper is presented at ICCV 2021.
https://ait.ethz.ch/projects/2021/PeCLR/
MIT License
83 stars 14 forks source link

Loss Weight to Train the RN_25D_wMLPref #9

Closed fylwen closed 2 years ago

fylwen commented 2 years ago

Hi, thanks for the released codes, and I have some questions related to the training of the 2.5D hand representation for hand pose estimation.

From the related papers, I consider that for the hand pose estimator the loss has three terms: 1) 2D pixel coordinate on the image plane, 2) the scale normalized relative z, and 3) the scale normalized z root (after refinement). I wonder whether my understanding is correct, and wonder about the value of weight parameters used to balance these three loss terms.

Thanks for your help!

spurra commented 2 years ago

Your understanding is correct! We use the same weights as explained in the supplementary of the following paper: https://arxiv.org/pdf/2003.09282.pdf (p.27)

i.e l_2d = 1, l_zrel = 5, l_zroot = 1