Does Catmull - Rom pass through the input samples?

The standard interpolating construction does at the sample positions.

Can it create values above the original maximum?

Yes. Negative kernel weights can cause overshoot.

Is every bicubic resize Catmull - Rom?

No. Bicubic methods can use different kernels and parameters.

AI Glossary / Computer vision and multimodal AI / Intermediate

Catmull - Rom

Definition

Catmull - Rom is a family of interpolating cubic constructions. In image resampling, the name commonly refers to a cubic filter that passes through the original samples and uses neighboring values, including negative weights, to produce a relatively sharp interpolated result.

What it means

Resampling estimates image values at positions that do not coincide with the original pixel grid. Nearest-neighbor methods copy a value, linear methods blend nearby values and cubic methods use a richer local curve. The standard uniform Catmull - Rom kernel is one such cubic choice.
Interpolating means that evaluating at an original sample position reproduces that sample under the intended construction. It does not mean that the values between samples are known to be correct.
Catmull - Rom also appears in curve design and animation. This entry follows the supplied photography source's resampling sense. Different spline parameterizations should not be assumed to have identical behavior to the uniform image-filter kernel.

How it works

For one-dimensional resampling, locate the desired position between two samples and use four neighboring samples to evaluate a cubic interpolant. The standard uniform Catmull - Rom form corresponds to a common cubic-convolution choice with parameter a=−0.5.
For images, a separable implementation applies the one-dimensional kernel along each axis. Negative side weights can preserve apparent sharpness but allow overshoot and undershoot near abrupt changes. Border extension and color representation affect the result. When reducing image size, appropriate anti-aliasing or scale-aware filtering is also needed; simply using an interpolating kernel at fewer output positions is not automatically sufficient to suppress aliasing.

Three practical examples

An image editor uses a Catmull - Rom-style cubic kernel when enlarging or transforming an image.
A rendering pipeline resamples a texture while balancing sharpness and ringing.
A graphics exercise compares cubic interpolation with linear interpolation on a known one-dimensional signal.

Think of it this way

Think of drawing a smooth curve through measured points while using neighboring points to guide its shape. The curve can look more natural than straight segments, but it may bend above or below the observed values. The analogy captures interpolation and overshoot; the particular kernel and parameterization determine the actual numerical behavior.

Worked example

For the standard uniform Catmull - Rom interpolant, the midpoint between the middle two of four equally spaced samples has weights [−1/16,9/16,9/16,−1/16]. These sum to one, so a constant signal remains constant.
Use sample values [0,10,10,0]. The midpoint value is 0+(9/16)10+(9/16)10+0=11.25. The interpolated value exceeds the largest input value, 10, because the outer weights are negative and the middle weights sum to more than one.
Linear interpolation between the middle values would return 10 instead. The cubic result may appear sharper in some image settings, but it can also create a halo or out-of-range value. This example uses the stated uniform kernel; other cubic filters and nonuniform Catmull - Rom spline parameterizations can behave differently.

Illustrative case study

Fictional case: Maple Graphics resizes high-contrast product labels with a sharp cubic filter. Thin characters look crisp, but bright outlines appear near dark edges. A developer initially treats the outlines as recovered detail.
The team compares interpolation kernels, checks whether output values overshoot the input range and evaluates readability at the final display size. It also reviews anti-aliasing for reductions rather than assuming the same settings suit every scale change. This case illustrates the sharpness - ringing trade-off, not a claim that Catmull - Rom is always preferable or always unsuitable for text imagery.

Common misconceptions and mistakes

Interpolating does not mean perfectly reconstructing an unknown scene. Passing through samples is a mathematical property, not evidence that intermediate values match reality. Negative weights can produce values outside the original range.
Catmull - Rom is not synonymous with every bicubic implementation. Libraries can use different cubic parameters or boundary rules. Enlargement and reduction have different sampling requirements. Clamping overshoots may change the interpolation result, and resampling encoded color values may differ from resampling linear-light values. State the implementation rather than relying on the filter name alone.

Business use

The kernel is relevant to image transformation, texture processing and visual-production pipelines where sharpness and computational simplicity matter. Its suitability depends on the content and final use.
Compare representative smooth gradients, fine textures and high-contrast edges. Inspect halos and aliasing at the actual output size, not only at a convenient zoom level. Record filter parameters, border policy and color-space handling for reproducibility. For quantitative images, verify that overshoot or clamping does not distort measurements. A visually sharper image can be less faithful in some respects, so evaluation should match the product's purpose.

Check your understanding

A resampled value exceeds every neighboring input value. Must the implementation be broken?

Answer: Not necessarily. A Catmull - Rom-style kernel can overshoot because some weights are negative. Whether that behavior is acceptable depends on the intended image-processing task.

Related terms

Resampling
Linear interpolation
Cubic interpolation
bicubic interpolation
interpolation kernel
ringing
overshoot
anti-aliasing

Learn next

Learn linear interpolation and cubic-convolution kernels next. Plot the kernel weights and apply them to a step edge and a constant signal. Then compare enlargement with downsampling to see why interpolation and anti-aliasing are related but distinct requirements.

Linear interpolation

Sources

Computational Photography glossary, Catmull–Rom and resampling context: https://people.csail.mit.edu/fredo/comp-photo-book/glossary.html Robert G. Keys, Cubic Convolution Interpolation for Digital Image Processing: https://doi.org/10.1109/TASSP.1981.1163711 Intel, Interpolation with Two-Parameter Cubic Filters: https://www.intel.com/content/www/us/en/docs/ipp/developer-guide-reference/2021-12/interpolation-with-two-parameter-cubic-filters.html Supplied kernel definition read and publication DOI located through bibliographic records. Full primary paper and final interpolation review remain pending. The midpoint calculation is original and specifies the uniform kernel.

Previous: Category utility Next: Cauchy-Schwarz inequality