MonocularRGB_2D_Handjoints_MVA19 icon indicating copy to clipboard operation
MonocularRGB_2D_Handjoints_MVA19 copied to clipboard

Published 20 hours ago • verified publisher icon FORTH-ModelBasedTracker

About model evaluation

Open halcyon370 opened this issue 4 years ago • 2 comments

Hello, is this model file "mobnet4f_cmu_adadelta_t1_model.pb" the same one just mentioned in your paper? I tried to evaluate on the CMU(MPII/NZSL) dataset, and I got AUC 0.875, which I expected to be 0.924 according to the paper. Could you give me some information about the evaluation details? Thanks!

halcyon370 avatar Apr 19 '20 11:04 halcyon370

The same problem.

www516717402 avatar May 07 '20 09:05 www516717402

Hello. In order to reproduce the results you have to crop around the center of the hand. As we mention in the paper:

For CMU we crop the images to a square box centered on the hand with side size of 1.2H, where H is the size of the person’s head.

Also, keep in mind that the model uploaded is trained only in the CMU dataset (in the paper it is denoted as CMU (224x224) ). Consequently, the expected performance on CMU should be 0.919 and not 0.924.

Best Regards.

philipposg avatar May 25 '20 11:05 philipposg