SplineOps

Projection-based antialiased resizing for N-D scientific arrays.

SplineOps is designed for precise, repeatable downsampling of regular-grid 2-D images and 3-D volumes. It combines explicit coordinate semantics, a readable Python reference implementation, a native CPU backend, and reusable plans for fixed-geometry workloads.

Installation

SplineOps requires Python 3.11 or newer:

python -m pip install splineops

Published wheels include the native resize extension on supported platforms.

Quick start

import numpy as np
from splineops import resize

volume = np.random.default_rng(0).random((64, 192, 192), dtype=np.float32)
coarse = resize(
    volume,
    output_size=(32, 96, 96),
    method="cubic-antialiasing",
)

The *-antialiasing methods project the input spline onto a coarser spline space instead of treating downsampling as interpolation alone. For repeated arrays with the same geometry, ResizePlan reuses setup and workspace state.

When to use SplineOps

SplineOps is a good fit when:

• continuous-valued 2-D or 3-D NumPy data must be downsampled;
• coordinate, boundary, and output-shape semantics must be explicit;
• aliasing matters; or
• many arrays share one resize geometry.

Another tool may be a better fit when:

• differentiable GPU execution is required;
• physical-space image metadata must be managed automatically;
• ordinary display-image scaling is the only goal; or
• categorical labels are being resampled, where nearest-neighbour semantics are normally appropriate.

Public maturity

Stability covers the documented CPU contracts and precision. CuPy remains experimental, and order-5 wavelets retain their approximate reconstruction contract. See API stability and support for the compatibility policy and exact boundaries.

Numerical and performance position

SplineOps does not claim to be the fastest generic 2-D image resizer. OpenCV, PyTorch, or other libraries can be faster when their coordinate, boundary, kernel, and antialiasing conventions are acceptable.

SplineOps is strongest when the spline model matters: defined grids and boundaries, N-D projection antialiasing, native/reference parity, and repeated fixed-geometry execution. Cross-library reports include semantic differences as well as timings; numerical difference from SplineOps is not treated as an independent accuracy metric.

The flagship SELMA3D microvessel validation uses 18 held-out expert-labelled 3-D microscopy patches. For the tested lateral 2x coarsening, SplineOps had higher mean vessel-ranking ROC AUC than the named SciPy, scikit-image, and PyTorch pipelines under a predeclared family-wise rule. It was 12.49x, 15.17x, and 1.36x faster, respectively, on the recorded one-thread CPU run. This is narrow preprocessing superiority—not segmentation or universal resampling superiority.

Explore that result interactively with the installed SELMA3D napari comparison demo. It downloads one checksum-verified public patch, runs the named methods locally, and provides a curtain slider over their outputs and expert vessel labels. The default demo extra avoids PyTorch; install selma3d-demo-all when the PyTorch-area comparison is required.

The exact approved public wording and its limits are maintained in Claims and evidence.

The 3-D microscopy case study tests that position on a public two-channel volume. It publishes the input checksum, code, raw results, local timings, and the limits of each metric.

The stricter BBBC050 segmentation validation uses manual labels and embryo-level held-out evaluation. SplineOps was fast and competitive, but the predeclared quality-superiority criterion was not met.

The controlled 3-D spectral-coarsening validation uses analytical targets rather than a library output. Its frozen criterion passed against six generic resize alternatives: projection reduced mean NRMSE by 54% versus the nearest frozen baseline over 72 smooth field cases. A post-hoc polyphase FIR audit was far more accurate, so this is not evidence of broad scientific-resampling or downstream-task superiority.

Backends

• NumPy is supported across the package.
• The native C++ backend accelerates resize on CPU.
• CuPy interoperability is experimental and limited to parts of TensorSpline.

Set SPLINEOPS_ACCEL=never to force the Python resize reference or SPLINEOPS_ACCEL=always to require the native extension. See the backend contract for exact scope.

Documentation

• Getting started
• Modules
• Examples
• Claims and evidence
• Resize guide
• Volume downsampling tutorial
• SELMA3D microvessel validation
• Interactive SELMA3D napari demo
• Controlled 3-D spectral-coarsening validation
• 3-D microscopy case study
• BBBC050 segmentation validation
• Performance evidence
• API reference
• Provenance
• Changelog

SplineOps modernizes spline methods developed across the Biomedical Imaging Group at EPFL and its collaborators. Method citations, implementation history, and source provenance are recorded explicitly.

Development

git clone https://github.com/splineops/splineops.git
cd splineops
python -m venv .venv
source .venv/bin/activate
python -m pip install -e '.[dev]'
python -m pytest -q

Reproducible resize evidence commands include:

python scripts/benchmark_resize_pr.py --profile smoke --output-dir /tmp/splineops-smoke
python scripts/benchmark_resize_native.py --backend both --output-csv /tmp/splineops-native.csv
python scripts/benchmark_resize_libraries.py --output-csv /tmp/splineops-libraries.csv