AutoSDF

Question 1

AudoSDF - What is the problem of the encoder working on the whole 3D shape and not on patches?

Answer 1

Each latent vector is seeing the whole shape which interferes with shape completion. We want a way to match partial view of a shape with partial view of latent codes.

Question 2

AudoSDF - What is the problem with transformers learning on 3D shapes?

Answer 2

The complexity of the attention mechanism goes up quadratically with regards to the dimension of the input.

Question 3

AudoSDF - How the complexity problem is being delt with?

Answer 3

They use VQ-VAE to learn the discrete latent representation of each 3D shape. Then the transformers trained are on an input with less dimensions.

Question 4

AudoSDF - What is the typical assumption of the ordering of the latent vectors? And how is it being used in relation to the distribution?

Answer 4

A raster scan ordering. Which then autoregressive models use to break down the distribution p(Z) = Π i=[1->d,1->d,1->d] p(zi|z<i).

Question 5

AudoSDF - What is the problem they are trying to solve when doing a rastering scan distribution break down?

Answer 5

For shape completion we don’t want to restrict ourselves to complete only the last tokens from the beginning. Most of the times the tokens being ‘seen’ are in a random oder.

Question 6

AudoSDF - How do they overcome the rastering scan order problem?

Answer 6

Assume that the joint distribution can broken down in terms of a random observable set of previous latent variables.

Question 7

AudoSDF - How do they model the prediction problem?

Answer 7

The distribution of the latent variable in a random place i is modelled by a transformer given all the previous observed variables.

Question 8

AudoSDF - What is the naive decomposition of P(Z|C)? (latent variable given a condition)

Answer 8

p(Z|C) = Πi p(zi |z<i , C)