Fitting code

[amundra15]

Hi,

Thanks for the amazing work. I want to try NIMBLE for my project. Do you have the optimization code for getting the NIMBLE parameters using the ground-truth mesh and texture map?

Also, can you provide the UV coordinates for your mesh? Right now I am using the MANO UV coordinates to generate my texture map, but that is not aligned with the NIMBLE-generated maps.

Best,
Akshay

[reyuwei]

Hi
I no longer have the optimization codes. But the structure should be similar to nr_reg here, just change the optimization variable to NIMBLE parameters, and use per-vertex distance as loss.

The UV coordinates is embedded in the model file. You can refer to this function to see how I export the geometry with uv coordinates.

Thanks!

[amundra15]

Hey
As a follow-up, do you have the RGB-based fitting code? I have multi-view images, and I want to use one/multiple of them to fit NIMBLE. It would be great if you could provide the optimization code you used in the paper, as re-implementing it from scratch could lead to sub-optimal performance.

[reyuwei]

The RGB-based result in our paper is from a learning based method (I2L-MeshNet) not fitting. If you have one or multiple images, I suggest first get the joint positions and then regressing the parameters according to joint positions. The optimization code should be quite similar to nr-reg.

[amundra15]

If I understand this correctly, this will only give you the geometry, right? How do you get the appearance?

[reyuwei]

You can use the photometric loss described in HTML. The process would be setting appearance parameter as an variable for optimization, then for each image, use a differentiable renderer (such as Pytorch3D) to compute photometric loss. You can set a fixed lighting condition for all views.