Streaming point cloud data (for online learning, not datasets)

[contactrika]

What is a good way to run methods from torch_geometric library on point clouds that come in dynamically (not on a fixed dataset)?
For example, we get data from a set of N environments in parallel and would like to process them right away (so streaming data comes in small batches, e.g. for online learning). We currently use custom code for taking a flat batch of point cloud observations - a tensor of size [num_envs x num_points_in_one_cloud x 3], then use a simple PointNet implementation to output a feature vector (and then use these lower-dimensional features in our active learning methods).
We would like to use the PointNet++, PCNN, DGCNN implementations from torch_geometric instead, but it seems inefficient to have to construct Data(pos=pts[env_i]) from each point cloud in the tensor into data_list and then aggregate them using Batch.from_data_list(data_list). All our point clouds are the same size. What would be a simple and better way to make data.x, data.pos, data.batch ?

Another question: what should data.x contain for point clouds? Several torch_geometric implementations assume data.x is present, but for point clouds with no additional metadata - should this be just same as data.pos or something else?

Would appreciate suggestions. Thanks!

[rusty1s]

A faster way to create Data objects and batching them together is:

batch = torch.arange(num_envs).view(-1, 1).repeat(1, num_points_in_one_cloud).view(-1)
data = Batch(pos=pos.view(-1, 3), batch=batch)

In case you do not have any node features, there are multiple options:

1. x=None, which is supported for, e.g., PointConv.
2. x=pos, which works quite well in practice, although you lose out on translation-invariance.
3. x=torch.ones(num_points, 1) for creating initial dummy node-features.