Modeling multiple shapes & differentiable renderer about nglod HOT 2 CLOSED

Thanks for your interest in our work!

1. We have, and you could consider the "generalization" experiment in the paper to be one form of this but with a variable size latent code.

2. This is more on the line of possible future work, but the code implemented as is can be used as a differentiable renderer.

from nglod.

Much appreciated, @tovacinni!

Can you kindly elaborate on 1.? My understanding of the "generalization" experiment under 4.3 is that you overfit sparse features + MLP weights to individual shapes before fixing the MLP weights and, again, overfitting sparse features to a second shape. Is my understanding correct? What I wonder here is whether one could "sample" shapes from nglod just as one can using DeepSDF (coming down to sampling from a mutivariate Gaussian) after training on a shape repository instead of an individual shape.

from nglod.