MultiViewGenerator#

class MultiViewGenerator(global_view_num=2, global_view_scale=(0.4, 1.0), local_view_num=4, local_view_scale=(0.1, 0.4), global_shared_transform=None, global_transform=None, local_transform=None, max_size=65536, center_height_scale=(0, 1), shared_global_view=False, center_sampling='random', center_sampling_kwargs=None, view_keys=('coord', 'origin_coord', 'color', 'normal'), anchor_bias_ratio=0.6, anchor_radius_scale=1.5, anchor_keys=('endpoints', 'branches_track', 'branches_shower', 'bragg'))[source]#

Bases: object

Generate multiple global and local crops for DINO/iBOT-style SSL.

Crops a major global view around a randomly chosen center (restricted in height by center_height_scale), then additional global views and several local views, each a nearest-k neighborhood whose size is a random fraction of the cloud (per global_view_scale / local_view_scale). Local centers are pushed to cover the major view and, when data_dict["anchors"] is available, a fraction (anchor_bias_ratio) of local crops are centered on anchors. Reads data_dict["coord"] and the keys in view_keys (and optional anchors); writes concatenated global_<key> / local_<key> arrays with global_offset / local_offset boundaries. Registered as MultiViewGenerator — use this string as the type in a transform=[...] config list.

Parameters:

  • global_view_num (int) – Number of global views (including the major one). Defaults to 2.

  • global_view_scale (tuple) – (min, max) fraction of points per global crop. Defaults to (0.4, 1.0).

  • local_view_num (int) – Number of local views. Defaults to 4.

  • local_view_scale (tuple) – (min, max) fraction of points per local crop. Defaults to (0.1, 0.4).

  • global_shared_transform (list, optional) – Transform config applied once to the whole sample before any cropping. Defaults to None.

  • global_transform (list, optional) – Transform config applied to each global view. Defaults to None.

  • local_transform (list, optional) – Transform config applied to each local view. Defaults to None.

  • max_size (int) – Upper bound on the points considered per view. Defaults to 65536.

  • center_height_scale (tuple) – (min, max) fractional z-band restricting where the major-view center is sampled. Defaults to (0, 1).

  • shared_global_view (bool) – If True, the extra global views are copies of the major view rather than freshly cropped. Defaults to False.

  • center_sampling (str) – Center-selection strategy, "random" or "cnms" (coverage non-max suppression). Defaults to "random".

  • center_sampling_kwargs (dict, optional) – Extra kwargs for cnms when center_sampling="cnms". Defaults to None.

  • view_keys (tuple) – Keys cropped into each view; must include "coord". Defaults to ("coord", "origin_coord", "color", "normal").

  • anchor_bias_ratio (float) – Fraction of local views centered on anchors (when anchors are present). Defaults to 0.6.

  • anchor_radius_scale (float) – Radius scale for anchor-centered local crops (size scales roughly cubically with it). Defaults to 1.5.

  • anchor_keys (tuple) – Anchor sub-keys pooled as candidate centers; led is always excluded. Defaults to ("endpoints", "branches_track", "branches_shower", "bragg").

Example

>>> import numpy as np
>>> np.random.seed(0)
>>> data = {"coord": np.random.rand(200, 3).astype("f4"),
...         "color": np.random.rand(200, 3).astype("f4"),
...         "index_valid_keys": ["coord", "color"]}
>>> out = MultiViewGenerator(global_view_num=2, local_view_num=4,
...                          view_keys=("coord", "color"))(data)
>>> out["global_coord"].shape, out["global_offset"]   # 2 global crops concatenated
((356, 3), array([170, 356]))
>>> out["local_coord"].shape, out["local_offset"]     # 4 local crops concatenated
((223, 3), array([ 66, 134, 155, 223]))
# exact sizes vary with the RNG; offsets mark per-view boundaries
get_center(coord, mask=None)[source]#
get_view(point, center, scale, size_override: int | None = None)[source]#