search

Omergral / Semantify

55 stars 2 forks source link

Custom obj files #1

Open sauradip opened 5 months ago

sauradip commented 5 months ago

Hi, Thanks for publishing the awesome piece of work.

I just want to know if we can test these mappers on other obj files which are not in canonical pose. Is it that we need to again train the mappers for that pose separately ?

Omergral commented 5 months ago

Hey there!

Actually, you don't need to re-train the mappers. However, the current implementation works with 3DMMs, therefore you do need to pass "body_pose" parameters for the model to be in a desired pose.

The mappers modify only the shape parameters (i.e. "betas" for SMPL / SMPL-X), so if you use the "--A_pose" argument in the sliders application, the pose will be set to A pose.

The "A_pose" is just a numpy array of the rotations for the model to be in an "A" pose and can be found under "assets/smplx/a_pose.npy", so if you have a specific set of rotations you want to use you can set it this way.

If you want to use multiple poses (i.e. multiple rotations) then more changes are needed.

sauradip commented 5 months ago

Thanks for the response and the ideas. It is indeed helpful.

I had one more question regarding the environment. I have been trying to setup the system for this code. But i cannot do so due to conflict with dependencies and also pytorch 3D failure while installation. Can you kindly recheck the environment and may be update the environment files in the repository.

Thanks in advance.

Omergral commented 5 months ago

Sure Will create an updated version of the environment in the upcoming days. Will update here

Omergral commented 5 months ago

Note that the environment was created with CUDA 11.1 on ubuntu 18.04 with cudnn 8.0.5 and python 3.8

It will take some time to change all the requirements and update the packages, we will work on it, but for now if you can manage using CUDA 11.1 the environment should work