search

TerenceCYJ / S2HAND

Model-based 3D Hand Reconstruction via Self-Supervised Learning, CVPR2021
104 stars 13 forks source link

The Projected 2D Joints seems collaps ? #4

Closed tlok666 closed 3 years ago

tlok666 commented 3 years ago

Hi, Thank you for sharing this work.

I try to reproduce the training process on FreiHAND dataset. So, I cloned the original code in this repo.

Only changed: 1、the 'model = nn.DataParallel(model.cuda())' in 'train.py' into 'model = model.cuda()', since return a dict with nn.DataParallel returns error. 2、"train_batch":32, into "train_batch":32, in 'SSL-e2e.json'

But, I find the 'Projected 2D Joints 𝑱𝒑𝒓' is collapsing into a point. As the figure2.1 show in the picture below. 0002_0001700

Is there anything else I need to pay attention to?

TerenceCYJ commented 3 years ago

Hi. The render-related loss may cause the collapse when the pose and texture are not accurate. I would recommend you to try the stage-wise training procedure or use stage-wise training at the start to initialize the end-to-end learning.

tlok666 commented 3 years ago

Hi. The render-related loss may cause the collapse when the pose and texture are not accurate. I would recommend you to try the stage-wise training procedure or use stage-wise training at the start to initialize the end-to-end learning.

Thanks. I tried your suggestion and it works. By the way. Can I ask, what is the PyTorch version on your computer?(I tried several versions, but none of them support return 'dict' under 'nn.DataParallel' mode)

TerenceCYJ commented 3 years ago

The code was developed with PyTorch 1.1.

tlok666 commented 3 years ago

The code was developed with PyTorch 1.1.

Thanks for your answers.