/g/ you are my last hope

/g/ you are my last hope. So I’m a phd student and the most important thing I’ve learned is that all papers are smoke and academia is a mafia of snake oil salesmen and directors only want to publish, even if it’s trash.

I’m trying to implement a SMPL model from scratch, a deep learning model that takes images of people (full body) and reconstructs the 3D mesh. Basically it just estimates a set of shape and pose parameters that have some meaning based on the image and with those parameters you can reconstruct the 3D mesh. Those parameters are just computed using good PCA from a set of registered meshes (standard meshes where the i-th vertex is the same for all instances, if in one mesh the first vertex is the tip of the nose, for all meshes the first vertex is the tip of the nose) and just computes the displacements of an average mesh (GOD I love PCA but it seems you need to use transformers for everything, good god).

The problem is that there are many available models that are ready to use for human bodies (the original goal of the article), but there is no open implementation TO TRAIN THIS SHIT FROM STARTING POINT for an arbitrary set of meshes other than human bodies (with all the necessary stuff like registered meshes, images, etc.). The guys assume they are open, but they are strictly closed about the actual implementation and training of this shit.

Can a guru send me somewhere? Nothing works here, thanks anons.