Questions regarding the paper Neural Scene Representation for Locomotion and the Minkowski Engine

[jeferal]

Hi, thank you very much for creating this engine to work with sparse tensors and for maintaining over the past years.

I am trying to reproduce the results in the paper Neural Scene Representation for Locomotion on Structured Terrain and I have a question regarding some design choices.

I want to know first how are the skip connections implemented with the Minkowski Engine in the paper. Are the implemented by concatenating features with ME.cat or are they implemented using Sparse Tensor Arithmetics (A+B)?
(or neither of these options?)

If the option is concatenating along feature dimension, what happens with the coordinates that are different between the tensors? Is it some sort of concatenating features of the cells that are the closest?

Thank you very much for your help in advance!

[jeferal]

Hi @chrischoy, this question is closer to the paper on Neural Scene Representation for Locomotion on Structured Terrain rather than mantaining this repository, I completely understand this is no longer mantained.

Regarding the spatio temporal U-Net shown in the paper, I understand the latent tensor is likely a fully occupied block of dimension 4x4x4x2. I am however failing to understand how to handle the temporal component in the decoder to produce an output of dimension 64x64x64x1. Several options I guess are: decoder starts from a slice with temporal index = to previous prediction so would upsample from 4x4x4x1. Or decoder upsamples both dimensions and output of the network is a slice with t index = to previous prediction.

Thank you very much for your help!

Please, let me know if this is not the right place to ask questions on this paper.