Segmentation Network for Sparse data, pos=pos, x=None, y=len(n_points)

[davodogster]

I'm struggling to find a network for my task even though my task is a common one.

I just want to do point cloud semantic segmentation. I don't have normals, and each point has a class label.
Preferably the network can generate graphs on the fly in the forward pass using:
edge_index = knn_graph(pos, k=K, batch=batch, loop=LOOP)

Is it true that I have to use messagePassing if each example has a different number of samples (points)? Sparse data - but the mean number of points per example / point cloud = 50,000.
It would be great to use pointTransformer or similar, but it doesn't use MessagePassing.

Is there any existing code/network I can use in the pyG documentation?
Thank you

[rusty1s]

Thanks for your interest. We have a few examples regarding segmentation of point clouds, e.g., here and here. Note that normals are definitely not required for these examples. PointConv can make use of x=None inside message passing. PointTransformerConv would require that you need to create dummy input features first, e.g., via torch.ones(data.num_nodes, 1).

All these examples work for varying number of points and make use of MessagePassing under the hood.

[davodogster]

Thanks @rusty1s , I think those two links are for the same script.

1. It appears Point _Transformer_Segmentation.py has builtin code to automatically create the dummy features:
   

2. When I feed my segmentation data to the model it throws this error:
   

.
.
.

My data:

Any advice would be greatly appreciated thank you

[rusty1s]

Fixed the link above. You need to initialize your model with in_channels=1, that is Net(1, …).