Move copy method to improve efficiency in RandLANet's transform method

[gaetanochiriaco]

Checklist

• [X] I have searched for similar issues.
• [X] I have tested with the latest development wheel.
• [X] I have checked the release documentation and the latest documentation (for main branch).

Proposed new feature or change

Current Behavior

The transform method of the RandLANet class is currently creating copies of the input data before sampling points. This significantly impacts performance, especially when dealing with large pointcloud files containing millions of points.

https://github.com/isl-org/Open3D-ML/blob/473592d6abf2492f906495f7eb949e52c44fe6ae/ml3d/torch/models/randlanet.py#L170-L185

https://github.com/isl-org/Open3D-ML/blob/473592d6abf2492f906495f7eb949e52c44fe6ae/ml3d/datasets/samplers/semseg_random.py#L51-L53

Proposed Change

Move the copy() operations to after the point sampling. The point sampler doesn't perform any in-place operations on the pc array, so it shouldn't have any side effects on the original pointcloud.

pc = data["point"]  # full pointcloud (N,3)
label = data["label"]
feat = data["feat"] if data["feat"] is not None else None
tree = data["search_tree"]

selected_idxs, center_point = self.trans_point_sampler(
    pc=pc,
    feat=feat,
    label=label,
    search_tree=tree,
    num_points=self.cfg.num_points,
    sampler=self.cfg.get("sampler", None),
)  # Points are sampled from the whole pointcloud (n_points,3)
pc_sub = pc[selected_idxs]
pc = pc_sub.copy()
label_sub = label[selected_idxs]
label = label_sub.copy()
if feat is not None:
    feat_sub = feat[selected_idxs]
    feat = feat_sub.copy()

random.shuffle(idxs)
return idxs, center_point

References

No response

Additional information

Performance Improvement

I have run performance tests on both the current and proposed implementations, running a single epoch with this configuration ml3d/configs/randlanet_toronto3d.yml .