fabyday / deformable-shape-completion-with-graph-convolutional-autoencoders-public Goto Github PK

This is a very valuable repo!
Great work reproducing the approach from the "Deformable Shape Completion with Graph Convolutional Autoencoders" paper.
I'm working on some related works about 3D shape learning and I'm curous to know if the results from the reproduced model comparable to the results reported in the original paper?

Hello. I am trying to build a similar architecture (as yours in the attached image) purely for 3d point cloud reconstruction. My main concern is the absence of a Graph Unpooling layer in the decoder.

Having done Global Mean Pooling in the encoder, let's say I get a vector of (batch_size, 128) and my original point cloud count is 5000. Would it be correct to up-sample by using nn.Linear(128,128*5000) and then reshaping into (batch_size, 5000,128) ? I am a bit new to Graph Networks, I am able to use Message Passing for node dimensions change and Graph Pooling techniques such as SAGPool for down-sampling, my main obstacle is the decoding part where I have to up-scale the latent vectors again using either de-convolution or up-sampling hence the reason I am asking. Thanks in advance !