Render TailorNet meshes on RGB images for synthetic training

[kristijanbartol]

Subtasks

• [x] Render both body and cloth meshes
• [x] Set color for cloth mesh (keep the body in a default color/human-skin color)

Detected problems

• Meshes rendering does not guarantee that the clothes mesh is on top of body, without holes, as shown in the example image

Solution (see #30)

To move clothes on top of the body, the interpenetration solver function can be used. However, it is non-trivial to do this update as the function is written in NumPy, as well as the underlying code that uses psbody package. Therefore, the whole synthetic data generation code needs to be rewritten in NumPy (and the original TailorNet's SMPL4Garment class can be used instead of TorchSMPL4Garment, which is addressed in a separate issue #30 .

[kristijanbartol]

Comments on the problem 1 (clothes not on top of the body):

• probably need to remove interpenetrations using tailornet_for_garmentor/utils/interpenetrations.py -> remove_interpenetrations_fast(), i.e., need to rewrite the function for pytorch
• actually, much more natural choice is to rewrite the whole synthetic data generation logic in numpy and stop using on-the-fly generation (otherwise, I would basically have to rewrite psbody package in torch)
• interpenetration removal is a more general problem that also applies to garment meshes, for example, pant vs. T-shirt. In this case, we first need to remove interpenetrations between pants and body, then between T-shirt and body, and finally between T-shirt-body and pant-body (see this issue).

Note: the issue concluded in the "Solution" part of the original comment.

[kristijanbartol]

Reopen: It turns out that mesh rendering and offline data pre-generation are not that related, so it's still relevant to have this issue opened separately. In particular, the problem is indeed in the rendering part, which might be solved by using a different rendering technique (check POSA!).

[kristijanbartol]

Closing: Using trimesh + pyrender (see renderers/surreal_renderer.py).