Render a ply mesh with texture using your render code

[jinnan-chen]

Hi, thanks for your great work!
I want to render a mesh.ply with texture in ply file, how to modify the render_data.py to add mode to the scene? Thanks!

[ShunyuanZheng]

Hi, thanks for your interest.

I think GPS-Gaussian is not suitable for joint rendering with the scene since the unprojected Gaussian points formulate merely a partial representation of humans. However, you can try to save the Gaussian points as in 3DGS and viewed in SIBER. Note that, scale needs a log function and opacity requires an inverse sigmoid function as here.

[jinnan-chen]

Hi, thanks for replying!
In fact, my question is not about GPS-Gaussian. It is about the data preparation to render multi-view images from original mesh.

[ShunyuanZheng]

Sorry for misunderstanding the question.

You can load the .ply file with trimesh and export the .obj file with a texture map. Then you can render them as the human scans in render_data.py. If you want to add both scans of human and scene, modify these lines as

    if len(renderer.scene.models) >= 2:
        renderer.modify_model(0, obj, texture)
        renderer.modify_model(1, scene_obj, scene_texture)
    else:
        renderer.add_model(obj, texture)
        renderer.add_model(scene_obj, scene_texture)

[jinnan-chen]

Thanks, but how to load the .ply file with trimesh and export the .obj file with a texture map using scripts for a lot of mesh.ply files?

[ShunyuanZheng]

If the material is included in .ply file, just load and save it with Trimesh. However, if it is not included in .ply file, I think you need some other tools to process the data. You can use MeshLab to check whether the material is included.

[jinnan-chen]

However, for 2K2K dataset, meshes cannot be processed in this way. But it seems these meshes are rendered the tested in the paper. I wonder how are they rendered using provided code?

[ShunyuanZheng]

Sorry, I have not actually used 2K2K data for training, so I did not notice it was not feasible to render them with the provided code. I think you can use THuman2.1 data as an alternative since it has been extended from 500 to 2500 models.

[jinnan-chen]

It seems not released yet.