How can I use Xconv to perform deconvolution.

[divyanshj16]

❓ Questions & Help

Question 1.

In the paper it says, the conv and deconv are same operations in pointcnn. However, I cannot make it to work. In the forward of Xconv. I tried passing the features from the latest layer in x and pos, batch from the layer I want to deconv to. Currently I am using (knn_interpolate + xconv) inplace of deconv using xconv. How can I make it work the proper way?

Question 2.

The image shows that the input will have 2048 point and after applying Xconv the num_points decrease to 768. But I have seen that the number of points stay same after XConv operation In your pointcnn test files you have explicitly used a fps downsampling method to do that. Can you please clarify?

[rusty1s]

Can you show me a minimal example to reproduce this? XConv shouldn't change the number of nodes. Are you sure that all inputs (x, pos, and batch) share the same number of nodes?

[divyanshj16]

You are right.

I tried passing the features from the latest layer in x and pos, batch from the layer I want to deconv to.

This is throws error.

The way I am changing the number of nodes is (knn_interpolate + Xconv)

[divyanshj16]

Here in: https://github.com/yangyanli/PointCNN/blob/master/pointcnn_seg/s3dis_x8_2048_fps.py There are deconv parameters.

After going at this line in this file, I saw that the deconv parameters are also passed to the XConv in some way.

@rusty1s Are you looking into this?

[divyanshj16]

Can you confirm if the torch_geometric implementation of xconv conforms with the architecture described in the paper?

[rusty1s]

The XConv should work as the architecture described in the paper, but it does not automatically reduce or increase the point size. You need to do this manually.

[divyanshj16]

Okay. Thanks for your help.

[rusty1s]

Nonetheless, it should be possible to include this to our current XConv module (similar to the PointConv module which is able to operate on bipartite graphs). I label this as a feature request :)

[divyanshj16]

I would like to contribute to it. Can you help figuring me out where exactly to put this?

[rusty1s]

Sure, and thank your very much :) Do you know about the bipartite message passing API used in PointConv? Here, we have two sets of nodes that exchange messages, and that is why pos is getting passed as a tuple with two entries. Semantically, this means that the first entry passes its messages to the second entry as described by edge_index. For providing the XConv operator with this functionality, this means that the knn computation should be deployed on two different points sets (pos[0] and pos[1]). Therefore, you should use the knn method instead of knn_graph to achieve this. Feel free to ask any questions if my explanations are not clear :)

[anewlearner]

Is there any update?