Skip to content

feat(core): add point-search, boundary-normal and vertebra-angle helpers - #141

Merged
Hendrik-code merged 2 commits into
mainfrom
feat/upstream-point-and-normal-helpers
Aug 28, 2026
Merged

feat(core): add point-search, boundary-normal and vertebra-angle helpers#141
Hendrik-code merged 2 commits into
mainfrom
feat/upstream-point-and-normal-helpers

Conversation

@Hendrik-code

@Hendrik-code Hendrik-code commented Aug 27, 2026

Copy link
Copy Markdown
Owner

Upstreams three helpers found in sibling repos under ongoing_projects/hendrik/, where each had been re-implemented locally several times over, and consolidates the vector helpers that were already duplicated inside TPTBox.

What's added

np_utils

  • np_index — row indices of an (N, D) point array that exactly equal a given point.
  • np_find_closest_point_index — nearest-point search. For integer (voxel) arrays it tries an exact match first, which is the common case when the query already sits on the grid, then falls back to a distance search.
  • np_compute_boundary_normals — per-voxel surface normals where one label touches another: 6-connectivity contact test, then the Sobel gradient of a Gaussian-smoothed mask. np_compute_surface has no normals, and calculate_pca_normal_np returns a single principal axis for a whole segmentation, so this filled a real gap.
  • np_unit_vector / np_angle_between — canonical implementations. spine/spinestats/angles.py and core/poi_fun/ray_casting.py each carried their own copy; both now alias these. angles.py keeps exporting unit_vector / angle_between under their historic names, so nothing downstream breaks.

core/poi_fun/vertebra_direction

  • get_vert_direction_angles — angle between each of a vertebra's Posterior/Inferior/Right direction vectors and the corresponding global axis. A healthy frame stays near-orthogonal; an implausibly large angle flags direction landmarks that are wrong.

What's refactored

Four sites hand-rolled cdist/argmin and now use the shared search:

File Was
pixel_based_point_finder.py (get_nearest_neighbor) cdist(p.T, locs) + .argmin()
pixel_based_point_finder.py (point-set projection loop) cdist_to_point + np.argmin
endplates.py ×2 np.linalg.norm(… , axis=1)

Also removed a dead commented-out duplicate in endplates.py. cdist_to_point itself is unchanged — it was already optimal.

Notes on the source material

Two things did not survive review, and are worth recording:

  • fast_cdist was not ported. Its upstream comment claims it is "much faster than scipy.cdist". It is 2–6× slower — scipy's is compiled C, while the einsum form materialises the (N, 3) difference array. Benchmarked at n=1k→1M across float and int dtypes. Only its exact-match fast path was adopted, reimplemented on top of cdist.
  • The normals point into the label, not outward. The source docstring says outward; the gradient of a smoothed mask points toward increasing density. Verified on a two-slab volume (mean normal [-1, 0, 0] with label 1 at low x) and documented accordingly.

A fourth candidate — a topological hole fill — was dropped entirely: np_fill_holes is already backed by fill_voids (compiled C++) and matches 6-connected binary_fill_holes on both a sealed and a leaky shell. The candidate was a torch conv3d loop iterated to fixpoint.

Similarly, only the angle triple from the vertebra-orientation candidate is new. Its R, corpus_com and rel_to_corpus are already get_vert_direction_matrix, a POI lookup, and get_vert_direction_PIR respectively — confirmed identical over 50 random orthonormal frames rather than assumed.

Verification

  • Full suite: 478 passed, 4 skipped. ruff check clean; the flake8 CI gate (E9,F63,F7,F82) reports 0.
  • 14 new tests across unit_tests/test_nputils.py and unit_tests/test_poi_ops.py, covering exact/multiple/no-hit lookup, brute-force agreement of the nearest-point search on int and float clouds, unit-length and direction of the normals, non-touching and diagonal-only-contact edge cases, and radian/degree consistency of the angle triple.
  • All four refactors proven equivalent to their previous behaviour over randomized inputs, including the column-vector input shape that get_nearest_neighbor's old np.expand_dims path accepted.

🤖 Generated with Claude Code

@Hendrik-code @claude
feat(core): add point-search, boundary-normal and vertebra-angle helpers
73369fa 
Upstreams three helpers that sibling repos had each re-implemented locally,
and removes the duplicate vector helpers that already existed inside TPTBox.

np_utils:
- np_index / np_find_closest_point_index: row lookup and nearest-point search
  over an (N, D) point array. The nearest-point search tries an exact match
  first for integer (voxel) arrays, then falls back to scipy's cdist.
- np_compute_boundary_normals: per-voxel surface normals where one label
  touches another, via a 6-connectivity contact test and the Sobel gradient
  of a Gaussian-smoothed mask. Complements np_compute_surface, which has no
  normals, and calculate_pca_normal_np, which returns a single axis for a
  whole segmentation.
- np_unit_vector / np_angle_between: canonical implementations. angles.py and
  ray_casting.py both carried their own copy; both now alias these, and
  angles.py keeps its historic public names.

poi_fun.vertebra_direction:
- get_vert_direction_angles: angle between each of a vertebra's P/I/R
  direction vectors and the corresponding global axis. Useful as a QC signal
  for direction landmarks that are not orthonormal.

Four existing call sites now use the shared point search instead of
hand-rolling cdist/argmin: pixel_based_point_finder (get_nearest_neighbor and
the point-set projection loop) and endplates (two norm(axis=1) distance
computations). All four were verified equivalent to their previous behaviour
over randomized inputs.

Note on the source material: the upstream fast_cdist was documented as "much
faster than scipy.cdist" but benchmarks 2-6x slower, since scipy's is compiled
C while the einsum form materialises the difference array. Only its
exact-match fast path was adopted, on top of cdist.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@github-actions

github-actions Bot commented Aug 27, 2026 •
edited
Loading

Copy link
Copy Markdown

Benchmark comparison

Speed (wall time per call)

baseline 439b142 vs head 439b142 · python 3.11.16 · 5 repeats + 1 warmup

Showing the 5 most-changed measurements per case; the rest are collapsed. Rows with a baseline below 3 ms are tagged (noise) — runner jitter on those swamps any real change.

ct_3d — shape (73, 47, 73)

Measurement baseline ms (median ±½·range) head ms (median ±½·range) Δ % p
nii_save 11.85 ±0.09 11.50 ±0.05 -2.9% 0.001
nii_filter_connected_components 4.32 ±0.04 4.21 ±0.05 -2.6% 0.017
nii_fill_holes 4.69 ±0.07 4.79 ±0.08 +2.1% 0.171
nii_load_img 11.98 ±0.09 11.73 ±0.18 -2.1% 0.204
poi_calc_centroids_nocrop 4.44 ±0.03 4.36 ±0.29 -1.6% 0.603
… 31 more measurements
Measurement baseline ms (median ±½·range) head ms (median ±½·range) Δ % p
nii_rescale_seg 3.14 ±0.05 3.10 ±0.03 -1.2% 0.051
nii_erode_msk 6.12 ±0.07 6.20 ±0.10 +1.2% 0.258
nii_dilate_msk 119.28 ±0.22 118.24 ±0.48 -0.9% 0.009
nii_resample_from_to 41.95 ±0.20 41.63 ±0.17 -0.8% 0.011
poi_calc_poi_from_subreg_vert 11.61 ±0.07 11.52 ±0.03 -0.7% 0.076
nii_rescale 40.02 ±0.10 39.77 ±0.17 -0.6% 0.028
metric_voxels 250463.00 ±0.00 250463.00 ±0.00 +0.0% —
metric_labels 3.00 ±0.00 3.00 ±0.00 +0.0% —
metric_foreground_pct 14.51 ±0.00 14.51 ±0.00 +0.0% —
nii_get_array 0.06 ±0.00 0.03 ±0.00 -49.8% (noise) 0.000
nii_set_dtype 0.54 ±0.04 0.39 ±0.04 -27.5% (noise) 0.000
poi_reorient 0.42 ±0.03 0.32 ±0.01 -23.2% (noise) 0.001
nii_reorient 0.87 ±0.05 0.69 ±0.03 -20.5% (noise) 0.000
nii_pad_to 0.69 ±0.02 0.56 ±0.02 -17.8% (noise) 0.000
nii_extract_label 0.43 ±0.03 0.36 ±0.05 -17.1% (noise) 0.172
nii_apply_crop 0.64 ±0.04 0.54 ±0.02 -14.7% (noise) 0.004
poi_map_labels 0.17 ±0.02 0.15 ±0.01 -14.5% (noise) 0.010
nii_load_seg 2.98 ±0.06 2.67 ±0.01 -10.4% (noise) 0.000
poi_save 0.42 ±0.03 0.39 ±0.01 -8.1% (noise) 0.029
poi_rescale 0.20 ±0.03 0.19 ±0.01 -6.7% (noise) 0.147
poi_local_to_global_arr 0.26 ±0.00 0.25 ±0.01 -5.4% (noise) 0.002
poi_to_global 0.19 ±0.01 0.18 ±0.03 -4.8% (noise) 0.992
poi_resample_from_to 0.57 ±0.01 0.54 ±0.01 -4.3% (noise) 0.004
nii_map_labels 0.84 ±0.01 0.87 ±0.08 +4.0% (noise) 0.276
nii_unique 0.62 ±0.03 0.60 ±0.02 -3.5% (noise) 0.086
nii_get_connected_components 2.09 ±0.02 2.02 ±0.04 -3.3% (noise) 0.020
nii_center_of_masses 1.86 ±0.01 1.81 ±0.01 -2.8% (noise) 0.000
nii_compute_crop 0.32 ±0.02 0.31 ±0.01 -2.3% (noise) 0.736
poi_calc_centroids 2.47 ±0.02 2.42 ±0.04 -1.7% (noise) 0.036
poi_load 0.86 ±0.02 0.87 ±0.01 +0.8% (noise) 0.506
nii_volumes 1.91 ±0.03 1.91 ±0.01 -0.3% (noise) 0.309

ct_2d — shape (73, 47, 1)

Measurement baseline ms (median ±½·range) head ms (median ±½·range) Δ % p
nii_rescale 5.56 ±0.16 5.34 ±0.04 -4.0% 0.094
nii_resample_from_to 7.42 ±0.17 7.13 ±0.12 -3.9% 0.029
nii_dilate_msk 13.72 ±0.06 13.79 ±0.04 +0.5% 0.104
metric_voxels 3431.00 ±0.00 3431.00 ±0.00 +0.0% —
metric_labels 3.00 ±0.00 3.00 ±0.00 +0.0% —
… 30 more measurements
Measurement baseline ms (median ±½·range) head ms (median ±½·range) Δ % p
metric_foreground_pct 37.60 ±0.00 37.60 ±0.00 +0.0% —
nii_compute_crop 0.17 ±0.01 0.15 ±0.01 -12.8% (noise) 0.105
poi_map_labels 0.24 ±0.01 0.22 ±0.01 -7.7% (noise) 0.006
poi_resample_from_to 0.76 ±0.03 0.71 ±0.03 -7.0% (noise) 0.018
poi_local_to_global_arr 0.31 ±0.01 0.29 ±0.00 -6.6% (noise) 0.007
nii_fill_holes 0.85 ±0.04 0.90 ±0.02 +6.3% (noise) 0.045
nii_set_dtype 0.44 ±0.01 0.46 ±0.01 +6.0% (noise) 0.003
nii_get_connected_components 0.71 ±0.04 0.75 ±0.04 +5.6% (noise) 0.077
nii_volumes 0.31 ±0.01 0.29 ±0.01 -5.5% (noise) 0.054
nii_center_of_masses 0.23 ±0.01 0.24 ±0.02 +4.7% (noise) 0.359
poi_to_global 0.27 ±0.01 0.26 ±0.01 -4.3% (noise) 0.626
nii_erode_msk 0.92 ±0.02 0.96 ±0.03 +4.2% (noise) 0.041
nii_load_seg 1.73 ±0.09 1.66 ±0.02 -3.8% (noise) 0.425
nii_unique 0.15 ±0.01 0.14 ±0.01 -3.5% (noise) 0.242
nii_pad_to 0.58 ±0.03 0.60 ±0.03 +3.0% (noise) 0.178
nii_get_array 0.02 ±0.00 0.02 ±0.00 -2.6% (noise) 0.276
nii_map_labels 0.54 ±0.01 0.52 ±0.02 -2.4% (noise) 0.090
poi_reorient 0.47 ±0.01 0.48 ±0.01 +2.4% (noise) 0.192
nii_filter_connected_components 0.87 ±0.02 0.89 ±0.02 +2.4% (noise) 0.014
nii_reorient 0.84 ±0.03 0.82 ±0.01 -1.8% (noise) 0.114
poi_save 0.55 ±0.01 0.54 ±0.02 -1.7% (noise) 0.119
nii_rescale_seg 0.82 ±0.02 0.80 ±0.01 -1.6% (noise) 0.260
nii_extract_label 0.45 ±0.03 0.46 ±0.02 +1.5% (noise) 0.725
nii_save 1.42 ±0.05 1.40 ±0.02 -1.4% (noise) 0.198
nii_apply_crop 0.72 ±0.02 0.71 ±0.02 -1.2% (noise) 0.458
poi_load 1.01 ±0.02 1.00 ±0.02 -0.8% (noise) 0.069
poi_calc_centroids_nocrop 1.13 ±0.02 1.12 ±0.01 -0.8% (noise) 0.389
nii_load_img 1.50 ±0.05 1.49 ±0.04 -0.6% (noise) 0.562
poi_calc_centroids 0.95 ±0.03 0.94 ±0.04 -0.2% (noise) 0.760
poi_rescale 0.29 ±0.01 0.29 ±0.01 -0.2% (noise) 0.183

mri_3d — shape (68, 52, 67)

Measurement baseline ms (median ±½·range) head ms (median ±½·range) Δ % p
nii_filter_connected_components 4.75 ±0.06 4.61 ±0.02 -3.0% 0.003
nii_load_img 9.72 ±0.06 9.55 ±0.05 -1.8% 0.003
nii_fill_holes 5.64 ±0.02 5.54 ±0.04 -1.7% 0.004
nii_erode_msk 7.74 ±0.05 7.61 ±0.15 -1.6% 0.737
poi_calc_centroids_nocrop 9.05 ±0.07 9.16 ±0.08 +1.2% 0.547
… 31 more measurements
Measurement baseline ms (median ±½·range) head ms (median ±½·range) Δ % p
nii_save 9.64 ±0.08 9.56 ±0.04 -0.9% 0.017
nii_resample_from_to 39.38 ±0.34 39.06 ±0.19 -0.8% 0.057
nii_dilate_msk 153.89 ±0.57 152.97 ±0.46 -0.6% 0.033
nii_rescale_seg 3.14 ±0.07 3.12 ±0.01 -0.6% 0.130
nii_rescale 37.43 ±0.13 37.34 ±0.18 -0.2% 0.757
poi_calc_poi_from_subreg_vert 10.90 ±0.03 10.90 ±0.07 -0.0% 0.595
metric_voxels 236912.00 ±0.00 236912.00 ±0.00 +0.0% —
metric_labels 8.00 ±0.00 8.00 ±0.00 +0.0% —
metric_foreground_pct 18.21 ±0.00 18.21 ±0.00 +0.0% —
nii_get_array 0.04 ±0.00 0.06 ±0.00 +35.5% (noise) 0.000
nii_set_dtype 0.53 ±0.02 0.59 ±0.05 +12.1% (noise) 0.041
poi_to_global 0.29 ±0.02 0.31 ±0.01 +6.7% (noise) 0.032
nii_extract_label 0.54 ±0.05 0.51 ±0.02 -6.1% (noise) 0.053
nii_get_connected_components 2.22 ±0.07 2.10 ±0.03 -5.5% (noise) 0.011
poi_rescale 0.33 ±0.01 0.31 ±0.02 -5.3% (noise) 0.145
nii_pad_to 0.81 ±0.03 0.77 ±0.01 -5.0% (noise) 0.071
poi_resample_from_to 0.77 ±0.02 0.81 ±0.02 +4.8% (noise) 0.007
nii_load_seg 2.98 ±0.10 2.84 ±0.02 -4.4% (noise) 0.055
poi_save 0.57 ±0.02 0.60 ±0.01 +4.4% (noise) 0.082
poi_reorient 0.53 ±0.03 0.52 ±0.01 -3.3% (noise) 0.436
nii_unique 0.66 ±0.01 0.64 ±0.01 -3.0% (noise) 0.037
poi_map_labels 0.26 ±0.01 0.26 ±0.00 +2.7% (noise) 0.307
nii_reorient 0.88 ±0.06 0.90 ±0.02 +2.2% (noise) 0.537
poi_calc_centroids 2.93 ±0.03 2.89 ±0.04 -1.5% (noise) 0.057
poi_local_to_global_arr 0.29 ±0.01 0.29 ±0.01 +1.5% (noise) 0.244
nii_map_labels 1.05 ±0.02 1.03 ±0.02 -1.5% (noise) 0.052
poi_load 2.10 ±0.04 2.12 ±0.03 +1.0% (noise) 0.842
nii_apply_crop 0.77 ±0.02 0.76 ±0.02 -0.5% (noise) 0.273
nii_center_of_masses 2.10 ±0.02 2.10 ±0.02 -0.2% (noise) 0.578
nii_volumes 2.04 ±0.02 2.05 ±0.02 +0.1% (noise) 0.284
nii_compute_crop 0.37 ±0.01 0.37 ±0.01 +0.0% (noise) 0.360

mri_2d — shape (68, 52, 1)

Measurement baseline ms (median ±½·range) head ms (median ±½·range) Δ % p
nii_rescale 5.81 ±0.07 5.41 ±0.05 -6.8% 0.000
nii_resample_from_to 7.35 ±0.19 7.21 ±0.12 -2.0% 0.258
nii_dilate_msk 18.27 ±0.06 18.16 ±0.17 -0.6% 0.876
metric_voxels 3536.00 ±0.00 3536.00 ±0.00 +0.0% —
metric_labels 7.00 ±0.00 7.00 ±0.00 +0.0% —
… 30 more measurements
Measurement baseline ms (median ±½·range) head ms (median ±½·range) Δ % p
metric_foreground_pct 50.17 ±0.00 50.17 ±0.00 +0.0% —
nii_rescale_seg 0.86 ±0.04 0.79 ±0.01 -8.6% (noise) 0.005
poi_to_global 0.32 ±0.01 0.29 ±0.02 -8.2% (noise) 0.024
nii_center_of_masses 0.44 ±0.01 0.40 ±0.01 -8.1% (noise) 0.001
poi_rescale 0.33 ±0.03 0.30 ±0.01 -7.1% (noise) 0.170
nii_erode_msk 1.28 ±0.04 1.20 ±0.03 -6.8% (noise) 0.006
nii_reorient 0.90 ±0.03 0.84 ±0.01 -6.7% (noise) 0.003
poi_calc_centroids_nocrop 1.45 ±0.03 1.35 ±0.04 -6.5% (noise) 0.017
nii_unique 0.15 ±0.01 0.14 ±0.01 -5.2% (noise) 0.177
nii_get_array 0.02 ±0.00 0.02 ±0.00 -4.7% (noise) 0.533
nii_fill_holes 1.09 ±0.02 1.04 ±0.04 -4.6% (noise) 0.085
poi_reorient 0.53 ±0.01 0.51 ±0.02 -4.4% (noise) 0.043
nii_set_dtype 0.46 ±0.04 0.44 ±0.02 -4.2% (noise) 0.480
nii_extract_label 0.46 ±0.02 0.44 ±0.01 -4.2% (noise) 0.050
nii_compute_crop 0.17 ±0.02 0.17 ±0.01 -3.8% (noise) 0.907
nii_pad_to 0.62 ±0.02 0.60 ±0.02 -3.6% (noise) 0.487
nii_volumes 0.49 ±0.01 0.47 ±0.01 -3.6% (noise) 0.048
nii_load_img 1.46 ±0.03 1.51 ±0.11 +3.3% (noise) 0.284
poi_save 0.60 ±0.02 0.62 ±0.02 +2.6% (noise) 0.322
poi_local_to_global_arr 0.30 ±0.01 0.30 ±0.03 +2.6% (noise) 0.313
nii_save 1.41 ±0.02 1.39 ±0.02 -1.7% (noise) 0.315
poi_load 1.91 ±0.02 1.88 ±0.01 -1.1% (noise) 0.073
nii_get_connected_components 0.69 ±0.03 0.68 ±0.03 -1.1% (noise) 0.495
poi_calc_centroids 1.15 ±0.03 1.14 ±0.01 -1.0% (noise) 0.835
nii_filter_connected_components 0.90 ±0.02 0.89 ±0.00 -0.9% (noise) 0.066
nii_map_labels 0.55 ±0.04 0.55 ±0.04 -0.9% (noise) 0.626
nii_apply_crop 0.72 ±0.02 0.72 ±0.03 -0.9% (noise) 0.841
poi_map_labels 0.26 ±0.01 0.26 ±0.01 +0.4% (noise) 0.583
poi_resample_from_to 0.79 ±0.01 0.79 ±0.01 -0.4% (noise) 0.514
nii_load_seg 1.65 ±0.06 1.64 ±0.02 -0.1% (noise) 0.887

Gate: a measurement fails when the baseline is ≥ 1 ms, the median grows by ≥ 50%, and Welch's t-test gives p < 0.05. metric_* rows are context only.

Memory (peak RSS growth per call)

baseline 439b142 vs head 439b142 · python 3.11.16 · 5 repeats + 1 warmup · sampler proc-status, isolation fork · measurement floor ≈ 0.90 MiB

Showing the 5 most-changed measurements per case; the rest are collapsed. Rows with a baseline below 3 MiB are tagged (noise) — runner jitter on those swamps any real change.

ct_3d — shape (73, 47, 73)

Measurement baseline MiB (median ±½·range) head MiB (median ±½·range) Δ % p
nii_load_img 3.52 ±0.00 3.57 ±0.00 +1.4% —
poi_resample_from_to 5.18 ±0.00 5.14 ±0.00 -0.8% —
nii_set_dtype 5.18 ±0.00 5.15 ±0.00 -0.7% —
nii_apply_crop 5.81 ±0.00 5.77 ±0.00 -0.7% —
poi_calc_centroids_nocrop 5.87 ±0.00 5.83 ±0.00 -0.7% —
… 31 more measurements
Measurement baseline MiB (median ±½·range) head MiB (median ±½·range) Δ % p
nii_extract_label 5.31 ±0.00 5.27 ±0.00 -0.7% —
nii_pad_to 5.37 ±0.00 5.33 ±0.00 -0.7% —
poi_calc_centroids 6.00 ±0.00 5.96 ±0.00 -0.7% —
nii_map_labels 5.43 ±0.00 5.40 ±0.00 -0.6% —
nii_resample_from_to 6.20 ±0.00 6.16 ±0.00 -0.6% —
nii_fill_holes 6.23 ±0.00 6.19 ±0.00 -0.6% —
nii_reorient 5.81 ±0.00 5.77 ±0.00 -0.6% —
nii_rescale_seg 5.89 ±0.00 5.85 ±0.00 -0.6% —
nii_rescale 6.07 ±0.00 6.04 ±0.00 -0.6% —
nii_filter_connected_components 6.81 ±0.00 6.77 ±0.00 -0.6% —
poi_calc_poi_from_subreg_vert 7.06 ±0.00 7.02 ±0.00 -0.6% 0.000
nii_get_connected_components 6.56 ±0.00 6.52 ±0.00 -0.5% —
poi_to_global 3.08 ±0.00 3.07 ±0.00 -0.5% —
nii_dilate_msk 5.86 ±0.00 5.84 ±0.00 -0.5% —
nii_erode_msk 5.88 ±0.00 5.85 ±0.00 -0.5% —
nii_load_seg 5.91 ±0.00 5.94 ±0.00 +0.5% —
poi_reorient 3.83 ±0.00 3.82 ±0.00 -0.4% —
nii_center_of_masses 3.27 ±0.00 3.26 ±0.00 -0.4% —
nii_volumes 3.52 ±0.00 3.51 ±0.00 -0.3% —
nii_save 3.71 ±0.00 3.70 ±0.00 -0.3% —
metric_voxels 250463.00 ±0.00 250463.00 ±0.00 +0.0% —
metric_labels 3.00 ±0.00 3.00 ±0.00 +0.0% —
metric_foreground_pct 14.51 ±0.00 14.51 ±0.00 +0.0% —
poi_rescale 2.61 ±0.00 2.59 ±0.00 -0.6% (noise) —
poi_local_to_global_arr 2.96 ±0.00 2.94 ±0.00 -0.5% (noise) —
nii_get_array 1.64 ±0.00 1.65 ±0.00 +0.5% (noise) —
nii_compute_crop 2.89 ±0.00 2.88 ±0.00 -0.4% (noise) —
nii_unique 2.96 ±0.00 2.95 ±0.00 -0.4% (noise) —
poi_save 1.40 ±0.00 1.41 ±0.00 +0.3% (noise) —
poi_load 1.61 ±0.00 1.62 ±0.00 +0.2% (noise) —
poi_map_labels 1.77 ±0.00 1.78 ±0.00 +0.2% (noise) —

ct_2d — shape (73, 47, 1)

Measurement baseline MiB (median ±½·range) head MiB (median ±½·range) Δ % p
nii_load_img 3.46 ±0.00 3.51 ±0.00 +1.4% —
poi_resample_from_to 5.12 ±0.00 5.08 ±0.00 -0.8% —
nii_set_dtype 5.12 ±0.00 5.08 ±0.00 -0.8% —
nii_extract_label 5.25 ±0.00 5.21 ±0.00 -0.7% —
nii_pad_to 5.30 ±0.00 5.27 ±0.00 -0.7% —
… 30 more measurements
Measurement baseline MiB (median ±½·range) head MiB (median ±½·range) Δ % p
nii_map_labels 5.37 ±0.00 5.33 ±0.00 -0.7% —
nii_fill_holes 6.16 ±0.00 6.12 ±0.00 -0.7% —
nii_reorient 5.75 ±0.00 5.71 ±0.00 -0.7% —
nii_apply_crop 5.75 ±0.00 5.71 ±0.00 -0.7% —
nii_rescale_seg 5.82 ±0.00 5.79 ±0.00 -0.7% —
poi_calc_centroids_nocrop 5.87 ±0.00 5.83 ±0.00 -0.7% —
poi_calc_centroids 5.93 ±0.00 5.89 ±0.00 -0.7% 0.000
nii_rescale 6.01 ±0.00 5.97 ±0.00 -0.6% —
nii_resample_from_to 6.14 ±0.00 6.10 ±0.00 -0.6% —
nii_get_connected_components 6.50 ±0.00 6.46 ±0.00 -0.6% —
nii_filter_connected_components 6.75 ±0.00 6.71 ±0.00 -0.6% —
nii_dilate_msk 5.80 ±0.00 5.77 ±0.00 -0.5% —
nii_erode_msk 5.82 ±0.00 5.79 ±0.00 -0.5% —
poi_to_global 3.02 ±0.00 3.00 ±0.00 -0.5% —
nii_center_of_masses 3.21 ±0.00 3.19 ±0.00 -0.5% —
nii_volumes 3.46 ±0.00 3.44 ±0.00 -0.5% —
nii_save 3.64 ±0.00 3.63 ±0.00 -0.4% —
poi_reorient 3.77 ±0.00 3.75 ±0.00 -0.4% —
nii_load_seg 5.85 ±0.00 5.88 ±0.00 +0.4% —
metric_voxels 3431.00 ±0.00 3431.00 ±0.00 +0.0% —
metric_labels 3.00 ±0.00 3.00 ±0.00 +0.0% —
metric_foreground_pct 37.60 ±0.00 37.60 ±0.00 +0.0% —
poi_rescale 2.54 ±0.00 2.52 ±0.00 -0.6% (noise) —
nii_compute_crop 2.83 ±0.00 2.82 ±0.00 -0.6% (noise) —
poi_local_to_global_arr 2.89 ±0.00 2.88 ±0.00 -0.5% (noise) —
nii_unique 2.89 ±0.00 2.88 ±0.00 -0.5% (noise) —
poi_save 1.34 ±0.00 1.34 ±0.00 +0.3% (noise) —
poi_load 1.55 ±0.00 1.55 ±0.00 +0.3% (noise) —
nii_get_array 1.58 ±0.00 1.59 ±0.00 +0.2% (noise) —
poi_map_labels 1.71 ±0.00 1.71 ±0.00 +0.2% (noise) —

mri_3d — shape (68, 52, 67)

Measurement baseline MiB (median ±½·range) head MiB (median ±½·range) Δ % p
nii_load_img 3.52 ±0.00 3.57 ±0.00 +1.3% —
poi_resample_from_to 5.12 ±0.00 5.08 ±0.00 -0.8% —
nii_set_dtype 5.18 ±0.00 5.14 ±0.00 -0.8% —
nii_extract_label 5.25 ±0.00 5.21 ±0.00 -0.7% —
nii_pad_to 5.30 ±0.00 5.27 ±0.00 -0.7% —
… 31 more measurements
Measurement baseline MiB (median ±½·range) head MiB (median ±½·range) Δ % p
nii_map_labels 5.37 ±0.00 5.33 ±0.00 -0.7% —
nii_fill_holes 6.16 ±0.00 6.12 ±0.00 -0.7% —
nii_reorient 5.75 ±0.00 5.71 ±0.00 -0.7% —
nii_apply_crop 5.75 ±0.00 5.71 ±0.00 -0.7% —
poi_calc_centroids_nocrop 5.80 ±0.00 5.77 ±0.00 -0.7% —
nii_rescale_seg 5.82 ±0.00 5.79 ±0.00 -0.7% —
poi_calc_centroids 5.93 ±0.00 5.89 ±0.00 -0.7% —
nii_rescale 6.01 ±0.00 5.97 ±0.00 -0.6% —
nii_resample_from_to 6.14 ±0.00 6.10 ±0.00 -0.6% —
nii_get_connected_components 6.50 ±0.00 6.46 ±0.00 -0.6% —
nii_filter_connected_components 6.75 ±0.00 6.71 ±0.00 -0.6% —
poi_calc_poi_from_subreg_vert 7.05 ±0.00 7.02 ±0.00 -0.6% —
nii_dilate_msk 5.80 ±0.00 5.77 ±0.00 -0.5% —
nii_erode_msk 5.82 ±0.00 5.79 ±0.00 -0.5% —
poi_to_global 3.02 ±0.00 3.00 ±0.00 -0.5% —
nii_center_of_masses 3.21 ±0.00 3.19 ±0.00 -0.5% —
nii_volumes 3.46 ±0.00 3.44 ±0.00 -0.5% —
nii_save 3.64 ±0.00 3.63 ±0.00 -0.4% —
poi_reorient 3.77 ±0.00 3.75 ±0.00 -0.4% —
nii_load_seg 5.85 ±0.00 5.88 ±0.00 +0.4% —
metric_voxels 236912.00 ±0.00 236912.00 ±0.00 +0.0% —
metric_labels 8.00 ±0.00 8.00 ±0.00 +0.0% —
metric_foreground_pct 18.21 ±0.00 18.21 ±0.00 +0.0% —
poi_rescale 2.54 ±0.00 2.52 ±0.00 -0.6% (noise) —
nii_compute_crop 2.83 ±0.00 2.82 ±0.00 -0.6% (noise) —
poi_local_to_global_arr 2.89 ±0.00 2.88 ±0.00 -0.5% (noise) —
nii_unique 2.89 ±0.00 2.88 ±0.00 -0.5% (noise) —
poi_save 1.40 ±0.00 1.40 ±0.00 +0.3% (noise) —
poi_load 1.55 ±0.00 1.55 ±0.00 +0.3% (noise) —
nii_get_array 1.58 ±0.00 1.59 ±0.00 +0.2% (noise) —
poi_map_labels 1.71 ±0.00 1.71 ±0.00 +0.2% (noise) —

mri_2d — shape (68, 52, 1)

Measurement baseline MiB (median ±½·range) head MiB (median ±½·range) Δ % p
nii_load_img 3.52 ±0.00 3.57 ±0.00 +1.3% —
poi_resample_from_to 5.12 ±0.00 5.08 ±0.00 -0.8% —
nii_set_dtype 5.18 ±0.00 5.14 ±0.00 -0.8% —
nii_extract_label 5.25 ±0.00 5.21 ±0.00 -0.7% —
nii_pad_to 5.30 ±0.00 5.27 ±0.00 -0.7% —
… 30 more measurements
Measurement baseline MiB (median ±½·range) head MiB (median ±½·range) Δ % p
nii_map_labels 5.37 ±0.00 5.33 ±0.00 -0.7% —
nii_fill_holes 6.16 ±0.00 6.12 ±0.00 -0.7% —
nii_reorient 5.75 ±0.00 5.71 ±0.00 -0.7% —
nii_apply_crop 5.75 ±0.00 5.71 ±0.00 -0.7% —
nii_rescale_seg 5.82 ±0.00 5.79 ±0.00 -0.7% —
poi_calc_centroids_nocrop 5.87 ±0.00 5.83 ±0.00 -0.7% —
poi_calc_centroids 5.93 ±0.00 5.89 ±0.00 -0.7% —
nii_rescale 6.01 ±0.00 5.97 ±0.00 -0.6% —
nii_resample_from_to 6.14 ±0.00 6.10 ±0.00 -0.6% —
nii_get_connected_components 6.50 ±0.00 6.46 ±0.00 -0.6% —
nii_filter_connected_components 6.75 ±0.00 6.71 ±0.00 -0.6% —
nii_dilate_msk 5.80 ±0.00 5.77 ±0.00 -0.5% —
nii_erode_msk 5.82 ±0.00 5.79 ±0.00 -0.5% —
poi_to_global 3.02 ±0.00 3.00 ±0.00 -0.5% —
nii_center_of_masses 3.21 ±0.00 3.19 ±0.00 -0.5% —
nii_volumes 3.46 ±0.00 3.44 ±0.00 -0.5% —
nii_save 3.64 ±0.00 3.63 ±0.00 -0.4% —
poi_reorient 3.77 ±0.00 3.75 ±0.00 -0.4% —
nii_load_seg 5.85 ±0.00 5.88 ±0.00 +0.4% —
metric_voxels 3536.00 ±0.00 3536.00 ±0.00 +0.0% —
metric_labels 7.00 ±0.00 7.00 ±0.00 +0.0% —
metric_foreground_pct 50.17 ±0.00 50.17 ±0.00 +0.0% —
poi_rescale 2.54 ±0.00 2.52 ±0.00 -0.6% (noise) —
nii_compute_crop 2.83 ±0.00 2.82 ±0.00 -0.6% (noise) —
poi_local_to_global_arr 2.89 ±0.00 2.88 ±0.00 -0.5% (noise) —
nii_unique 2.89 ±0.00 2.88 ±0.00 -0.5% (noise) —
poi_save 1.40 ±0.00 1.40 ±0.00 +0.3% (noise) —
poi_load 1.55 ±0.00 1.55 ±0.00 +0.3% (noise) —
nii_get_array 1.58 ±0.00 1.59 ±0.00 +0.2% (noise) —
poi_map_labels 1.71 ±0.00 1.71 ±0.00 +0.2% (noise) —

Gate: a measurement fails when the baseline is ≥ 1 MiB, the median grows by ≥ 50%, and Welch's t-test gives p < 0.05. metric_* rows are context only.

@Hendrik-code Hendrik-code self-assigned this Aug 27, 2026
@Hendrik-code Hendrik-code mentioned this pull request Aug 27, 2026
Merged
@Hendrik-code Hendrik-code added the enhancement New feature or request label Aug 28, 2026
@Hendrik-code
Hendrik-code merged commit 748c8e0 into main Aug 28, 2026
6 checks passed
@Hendrik-code
Hendrik-code deleted the feat/upstream-point-and-normal-helpers branch August 28, 2026 10:39
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@robert-graf robert-graf robert-graf approved these changes

Assignees

@Hendrik-code Hendrik-code

Labels

enhancement New feature or request

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@Hendrik-code @robert-graf