search

zijinxuxu / SMHR

Single-stage Multiple Hand Reconstruction
MIT License
29 stars 4 forks source link

Inquiry about poor results when producing demo on HO3D images #3

Closed viridityzhu closed 1 year ago

viridityzhu commented 1 year ago

Hi there,

First of all, I want to express my appreciation for the excellent work you have done!

When producing a demo on HO3D images, I follow your readme to modify the demo.sh as:

task = 'simplified',dataset='HO3D',Input_RES=224,CHECKPOINTS_PATH='${ROOT}/data/HO3D-2d-simplified.pth', 
comment out #L35 '--pick_hand'

where the HO3D-2d-simplified.pth is downloaded from the provided link.

But the results look poor for all test images. The predicted hand is always in the left-up corner: image

I noticed the warnings that may be causing the problem:

loaded HO3D-2d-simplified.pth
Drop parameter mean_theta.
Skip loading parameter p3.weight, required shapetorch.Size([512, 256, 4, 4]), loaded shapetorch.Size([256, 512, 3, 3]).
Skip loading parameter p5.weight, required shapetorch.Size([2048, 256, 8, 8]), loaded shapetorch.Size([2048, 256, 4, 4]).

I would greatly appreciate your assistance in resolving this issue.

Note:

  1. I've also tried to modify the input resolution, but it did not solve the problem
  2. I tried FreiHAND images using Single_pho.pth, and everything looked good.
  3. I also tried HO3D images using Single_pho.pth, and the results were equally poor: image
zijinxuxu commented 1 year ago

Hello, the models stored in the cloud disk might be from previous versions of the code. During the paper revision period, we have restructured the entire code framework. I suggest that you can train a model on your own using the HO3D dataset, and the training code should work fine.