vol_tiled_map — 3D domain op

Data kinds: voxelvoxel

Call: import fullseye as fs; fs.ledger.vol_tiled_map(vol, fn, tile=64, overlap=8) (to call the implementation directly, import volops; volops.vol_tiled_map(vol, fn, tile=64, overlap=8); from the registry, ops3d.get("vol_tiled_map"))

Usage

Apply a shape-preserving volume operator in overlapping z-slabs, so peak

working memory is bounded by the slab — not the volume.

The third leg of the memory family: :func:vol_crop_domain shrinks *where*

you compute, :mod:volregion shrinks *what you keep*, and this bounds *how

much lives in RAM at once*. Each slab `vol[z0-overlap : z1+overlap]` is

run through *fn* and only the core `[z0:z1)` of the result is kept, so a

volume far above an operator's comfortable size streams through in

constant-memory pieces (measured: peak working set of a Gaussian drops with

the slab size while the output stays exact — see the test).

Correctness contract (the honest part): the result equals `fn(vol)`

exactly only for *local* operators whose spatial footprint along z is at

most *overlap* voxels on each side (a Gaussian of sigma s with scipy's

default truncation needs `overlap >= round(4 * s)`; a morphology with a

k-voxel structuring element needs `overlap >= k`). A *global* operator

(Otsu, normalisation, anything that looks at the whole histogram) is

silently WRONG under tiling — this function cannot detect that, so it is

documented instead: do not tile global operators.

Parameters: *fn* is any callable mapping a `(d, H, W)` float64 volume to

an array of the same shape (e.g. `lambda v: vol_gradient_magnitude(v)`).

*tile* is the slab thickness (>= 1), *overlap* the per-side context

(>= 0). A slab result of the wrong shape raises `ValueError` immediately

(fail-closed — a shape-changing fn would silently corrupt the assembly).

References (sample data, literature)

• Sample-data catalog (download URLs / licences) — 2-D uses skimage.data (BSD/public domain) plus synthetic images; 3-D lists download URLs for real data sources (Stanford, PDS, …).

• Operator provenance and references — the sources of the research/methods this op family came from.

Runnable examples (verified samples that actually call this op)

rle_region_efficiencypy -3.11 examples_3d/rle_region_efficiency.py

Ops the type connects to (they accept voxel as input)

voxel_to_mips · voxel_to_mesh · signed_distance_field · to_points · sobel3d · hessian3d · curvature_maps · edt_jfa

Same category (domain)

vol_reduce_domain · vol_bounding_box · vol_crop_domain · vol_uncrop


*Provenance: volops.py — 3D operator registry. This per-op note is generated by tools/opdocs.py md (do not hand-edit).*

© 2026 Kazufumi Furuse — Fullseye operator documentation. Licensed under Apache-2.0.