search

garvita-tiwari / PoseNDF

Implementation of Pose-NDF: Modeling Human Pose Manifolds with Neural Distance Fields
Other
209 stars 19 forks source link

How to train a smooth manifold of poses? #26

Open HospitableHost opened 1 year ago

HospitableHost commented 1 year ago

Hi there,

I'm experiencing an issue about training a smooth manifold.

With your pretrained model, I find that the distances are always zero or very small, which means that your manifold is smooth and your pretrained model can always predict zero distances for plausible poses.

But with my model (which is trained by your codes for training and generating training data), I find that my manifold is not smooth, my model will predict zero distances or small distances for plausible poses, and predict large distances for unplausible poses.

In a word, the variance of distances from my model is larger than yours.

Could you please help me with this problem? I'd really appreciate any assistance you can provide.

Thanks in advance!

garvita-tiwari commented 1 year ago

My model will predict zero distances or small distances for plausible poses, and predict large distances for unplausible poses

Isn't this the desired behavior? Do you use AMASS_clean, AMASS_noisy and random poses during training?

Can you share a plot of GT distance v/s predicted distance? As long as it's linear, it should be a good model. and we can compare this with the old model.

garvita-tiwari commented 2 months ago

My model will predict zero distances or small distances for plausible poses, and predict large distances for unplausible poses

Isn't this the desired behavior? Do you use AMASS_clean, AMASS_noisy and random poses during training? Can you share a plot of GT distance v/s predicted distance? As long as it's linear, it should be a good model. and we can compare this with the old model.

Did your model use random poses for training? May I ask which part of the code is doing this?

We prepare random poses and their distance from NN in AMASS dataset offline and use it for training. We use this script to prepare this data: https://github.com/garvita-tiwari/PoseNDF/blob/main/data/prepare_traindata.py

ZhengdiYu commented 2 months ago

My model will predict zero distances or small distances for plausible poses, and predict large distances for unplausible poses

Isn't this the desired behavior? Do you use AMASS_clean, AMASS_noisy and random poses during training? Can you share a plot of GT distance v/s predicted distance? As long as it's linear, it should be a good model. and we can compare this with the old model.

Did your model use random poses for training? May I ask which part of the code is doing this?

We prepare random poses and their distance from NN in AMASS dataset offline and use it for training. We use this script to prepare this data: https://github.com/garvita-tiwari/PoseNDF/blob/main/data/prepare_traindata.py

Thank you for your reply. However, I think you are adding noise to the original AMASS poses, instead of directly generating random poses?