Caenorst
commented
2 years ago Hi @3a1b2c3 , Good question, we actually haven't designed an optimized function for that (but there is a good chance we may in the future as we can see some use cases), an easy way to do it would be to take the points at the highest level from the point_hierarchy, and do something similar to https://github.com/NVIDIAGameWorks/kaolin/blob/9085d1454c041b6300afc95ab6320e729430bb81/kaolin/ops/conversions/pointcloud.py#L171-L197
For a given point_hierarchy1 / point_hierarchy2 and pyramid1 / pyramid2 with matching features1 / features2 on the last level It would look something like that:
points1 = point_hierarchy1[pyramid1[-1, 1]:pyramid1[-1, 1] + pyramid1[-1, 0]]
points2 = point_hierarchy2[pyramid2[-1, 1]:pyramid2[-1, 1] + pyramid2[-1, 0]]
all_points = torch.cat([points_hierarchy1, points_hierarchy2], dim=0)
unique, unique_keys, unique_counts = torch.unique(all_points.contiguous(), dim=0,
return_inverse=True, return_counts=True)
morton, keys = torch.sort(points_to_morton(unique.contiguous()).contiguous())
points = morton_to_points(morton.contiguous())
merged_octree = unbatched_points_to_octree(points, level, sorted=True)
all_features = torch.cat([features1, features2], dim=0)
feat = torch.zeros(unique.shape[0], all_features.shape[1], device=all_features.device).double()
# Here we just do an average when both octrees have features on the same coordinate
feat = feat.index_add_(0, unique_keys, all_features.double()) / unique_counts[..., None].double()
feat = feat.to(all_features.dtype)
merged_features = feat[keys]If you are struggling with the variables I'm talking about, I suggest having a look at our SPC documentation to get more familiar with it.
3a1b2c3
I have a case where i want to merge 2 separate octrees, just wondering if this already possible?