Blur Shift: Concepts in Imaging, Learning & Dynamics

• Blur Shift is a family of concepts that captures spatial variability in motion blur, distribution shifts in learning systems, and symbolic dynamics compactification.
• In imaging, it commonly refers to the local PSF first moment, detailing motion-induced changes in centroid, orientation, and extent under varying conditions.
• In recognition and tracking, blur shift represents a domain mismatch between training and test conditions, impacting model accuracy unless compensatory strategies are applied.

Searching arXiv for the cited papers to ground the article. Blur shift is a non-universal technical term whose meaning depends on disciplinary context. In the cited literature, it denotes, most often, either a property of motion blur and point-spread functions (PSFs), a blur-induced distribution shift in learning systems, or a compactification construction in symbolic dynamics. This suggests that the term is best understood as a family of concepts organized around blur, displacement, and loss of shift-invariance rather than as a single standardized definition (Gast et al., 2016, Guo et al., 2019, Almeida et al., 2020).

1. Terminological scope

The cited literature uses “blur shift” in at least three distinct senses.

Domain	Meaning of “blur shift”	Representative papers
Motion blur and image formation	The centroid or first moment of a local PSF; more broadly, depth- or motion-dependent spatial variation of blur	(Gast et al., 2016, Torres et al., 2023, Lee et al., 2023)
Recognition, tracking, deblurring	A blur-induced train–test or benchmark mismatch in image statistics	(Vasiljevic et al., 2016, Guo et al., 2019, Wu et al., 2023)
Symbolic dynamics	A blur-symbol compactification of countable alphabet shift spaces	(Almeida et al., 2020, Garibaldi et al., 24 Jul 2025)

In imaging, the central issue is usually that blur is not globally shift-invariant. Localized object motion, parallax camera motion, atmospheric turbulence, focal-spot blur in CT, and synthetic dual-pixel view synthesis all produce PSFs whose orientation, extent, and sometimes centroid depend on position, depth, view, or exposure time (Chan, 2022, II et al., 2016, Abuolaim et al., 2021). In learning systems, by contrast, blur shift refers to a domain shift: models trained on sharp or weakly blurred data encounter altered frequency content and edge statistics at test time, with predictable degradation unless robustness is learned or augmentation expands the blur distribution (Vasiljevic et al., 2016, Wu et al., 2023).

A common misconception is that these usages are interchangeable. They are not. In one line of work, blur shift is a geometric quantity attached to a PSF; in another, it is a statistical mismatch between datasets; in symbolic dynamics, it is a topological construction based on adding blur symbols to represent infinite subsets of an alphabet (Garibaldi et al., 24 Jul 2025).

2. Blur shift as a PSF moment and as loss of shift-invariance

For localized object motion blur, the observed image is modeled by an exposure-time integral. With a binary mask $M(x) \in \{0,1\}$, a blurred image can be written as

$I_b(x) = (1 - M(x)) I_{bg}(x) + M(x) I_{obj,b}(x),$

with

$$I_{obj,b}(x) = \frac{1}{T} \int_0^T I_0(W_t(x))\,dt,$$

or, equivalently, in spatially varying convolution form,

$$I_{obj,b}(x) = \int K_x(\delta)\,I_0(x+\delta)\,d\delta.$$

In this formulation, “blur shift” refers to the first moment of the local PSF, namely its centroid, together with the associated orientation and extent. The centroid $\mu(x)$ encodes average displacement, orientation encodes motion direction, and extent encodes motion magnitude integrated over exposure (Gast et al., 2016).

For constant translational motion with velocity $v$ during exposure $T$, the PSF is a line kernel whose length equals $T\|v\|$ and whose orientation aligns with $v/\|v\|$. Its first moment is

$$\mu(x) = \int \delta\,K_x(\delta)\,d\delta \approx \frac{1}{T}\int_0^T (W_t(x)-x)\,dt,$$

which reduces to

$I_b(x) = (1 - M(x)) I_{bg}(x) + M(x) I_{obj,b}(x),$0

when the latent reference is taken at the start of exposure. If the latent image is defined at mid-exposure, the PSF is centered and $I_b(x) = (1 - M(x)) I_{bg}(x) + M(x) I_{obj,b}(x),$1 (Gast et al., 2016). This dependence on reference time is important: blur shift is not an invariant scalar independent of temporal convention.

Magnitude-only blind deconvolution introduces a further subtlety. In Lane-and-Bates-style cepstral analysis, a spatial shift of the kernel changes only the phase of the transfer function, so the absolute PSF centroid is not identifiable from $I_b(x) = (1 - M(x)) I_{bg}(x) + M(x) I_{obj,b}(x),$2. In that setting, the support or orientation of the PSF may be recoverable, but its absolute spatial shift is not; practical methods therefore impose a centered, zero-mean PSF [0609165]. This objective fact explains why “blur shift” is sometimes treated as a measurable first moment and sometimes as an unidentifiable nuisance, depending on whether phase information is modeled.

3. Motion from blur in a single image

“Parametric Object Motion from Blur” formulates motion blur not as a nuisance but as a signal for inferring object motion from a single image. The method combines a parametric object motion model with a segmentation mask, uses a differentiable image-formation model with respect to motion parameters, and generalizes marginal-likelihood techniques from uniform blind deblurring to localized, non-uniform blur. The proposed pipeline has two stages—first in derivative space and then in image space—and estimates both parametric object motion and motion segmentation from a single image alone (Gast et al., 2016).

A standard formulation consistent with this model uses an affine velocity field

$I_b(x) = (1 - M(x)) I_{bg}(x) + M(x) I_{obj,b}(x),$3

or a reduced rigid-with-scale parameterization, and couples it to a latent-image prior and a segmentation regularizer through an energy of the form

$I_b(x) = (1 - M(x)) I_{bg}(x) + M(x) I_{obj,b}(x),$4

Alternating or EM-style updates then optimize the latent image, motion parameters, and mask. The derivative-space stage reduces sensitivity to low-frequency content; the image-space stage refines the estimate at full resolution. A plausible implication is that blur shift, as encoded by local PSF moments, is most useful when embedded in a joint model rather than measured in isolation (Gast et al., 2016).

A non-parametric alternative appears in “From Motion Blur to Motion Flow,” where heterogeneous motion blur is treated explicitly as loss of shift-invariance. Each pixel has its own local PSF parameterized by a dense motion field $I_b(x) = (1 - M(x)) I_{bg}(x) + M(x) I_{obj,b}(x),$5, and a fully convolutional network estimates this dense motion flow directly from a blurred image. The local PSF is modeled as a unit-mass line segment aligned with $I_b(x) = (1 - M(x)) I_{bg}(x) + M(x) I_{obj,b}(x),$6, and deblurring then becomes nonblind inversion with spatially varying kernels (Gong et al., 2016). Relative to low-dimensional parametric models, this replaces segmentation-coupled motion estimation by dense per-pixel blur geometry.

4. Depth, camera motion, turbulence, and other physical forms of blur shift

Under parallax camera motion, blur shift becomes depth-dependent. In the Image Compositing Blur model, a point at depth $I_b(x) = (1 - M(x)) I_{bg}(x) + M(x) I_{obj,b}(x),$7 undergoes image displacement inversely proportional to depth, and the effective kernel length obeys the rule of thumb

$I_b(x) = (1 - M(x)) I_{bg}(x) + M(x) I_{obj,b}(x),$8

Near objects therefore produce longer blur kernels than far objects for the same camera trajectory. The model approximates the exposure integral by depth layers, per-layer kernels, and alpha compositing, which yields a computationally efficient forward model for spatially varying blur (Torres et al., 2023).

Radiance-field formulations make the same point in a stronger form: blur from 6-DOF camera motion is not a 2D convolution. In ExBluRF, a blurred pixel accumulates radiance along a time-varying ray determined by a camera trajectory in $I_b(x) = (1 - M(x)) I_{bg}(x) + M(x) I_{obj,b}(x),$9,

$$I_{obj,b}(x) = \frac{1}{T} \int_0^T I_0(W_t(x))\,dt,$$0

and the method jointly optimizes voxel-based radiance fields and Bézier-parameterized trajectories in blurred image space. The paper reports that this design delivers “the order of 10 times less training time and GPU memory consumption” than existing methods while improving sharpness on extremely blurred multi-view data (Lee et al., 2023). This is a direct consequence of treating blur shift as a three-dimensional geometric effect rather than a planar kernel.

Atmospheric turbulence raises a related ordering controversy. Decomposing pupil phase into tilt and higher-order aberrations yields a shifted PSF center from the tilt term and blur from the remaining aberrations. “Tilt-then-Blur or Blur-then-Tilt?” argues that the correct physical forward model is tilt-then-blur, not blur-then-tilt: the pupil-plane linear phase shifts the PSF in the image plane by $$I_{obj,b}(x) = \frac{1}{T} \int_0^T I_0(W_t(x))\,dt,$$1, and higher-order terms blur around that shifted center (Chan, 2022). For spatially invariant blur and a global shift, the two operators commute algebraically; the optical attribution nevertheless remains tilt first, blur second.

In CT and CBCT, blur shift denotes shift-variant focal-spot blur. Because the x-ray source is extended and the anode is angled, the apparent focal-spot length and orientation vary with detector position, ray angle, magnification, depth, and tube orientation. Both “Modeling shift-variant X-ray focal spot blur for high-resolution flat-panel cone-beam CT” and “High-Fidelity Modeling of Shift-Variant Focal-Spot Blur for High-Resolution CT” show that model-based iterative reconstruction with a shift-variant blur model outperforms reconstructions with no blur model or a single shift-invariant blur model, and that tube orientation changes the spatial distribution of “sweet spots” and degraded regions in the field of view (II et al., 2016, II et al., 2017).

5. Blur shift as a distribution shift in recognition, tracking, and deblurring

In recognition and tracking, blur shift refers to a mismatch between the blur statistics seen during training and those encountered at test time. “Examining the Impact of Blur on Recognition by Convolutional Networks” shows that CNNs trained only on high-quality images suffer marked degradation on blurred inputs, and that the degradation is attributable to a mismatch in low-level image statistics. On ImageNet with VGG-16 at scale 256, Top-5 accuracy drops from 90.88% on sharp images to 81.48% under defocus blur with radius $$I_{obj,b}(x) = \frac{1}{T} \int_0^T I_0(W_t(x))\,dt,$$2 and to 60.97% under $$I_{obj,b}(x) = \frac{1}{T} \int_0^T I_0(W_t(x))\,dt,$$3; fine-tuning on a sharp-plus-defocus mixture raises the $$I_{obj,b}(x) = \frac{1}{T} \int_0^T I_0(W_t(x))\,dt,$$4 result to 87.01% while leaving sharp-image accuracy nearly unchanged at 90.59% (Vasiljevic et al., 2016). The same study reports that blur robustness emerges mainly in deeper layers, where representations become more invariant.

Tracking work reaches a more qualified conclusion. On the Blurred Video Tracking benchmark, light blur can improve many trackers, but heavy blur always hurts overall robustness, and deblurring is not uniformly beneficial. The benchmark constructs blur by temporal averaging,

$$I_{obj,b}(x) = \frac{1}{T} \int_0^T I_0(W_t(x))\,dt,$$5

with blur levels $$I_{obj,b}(x) = \frac{1}{T} \int_0^T I_0(W_t(x))\,dt,$$6. The study reports that 17 of 23 trackers improve at $$I_{obj,b}(x) = \frac{1}{T} \int_0^T I_0(W_t(x))\,dt,$$7 and 14 of 23 at $$I_{obj,b}(x) = \frac{1}{T} \int_0^T I_0(W_t(x))\,dt,$$8, whereas only 7 improve at $$I_{obj,b}(x) = \frac{1}{T} \int_0^T I_0(W_t(x))\,dt,$$9 and 2 at $$I_{obj,b}(x) = \int K_x(\delta)\,I_0(x+\delta)\,d\delta.$$0; full deblurring may help heavily blurred videos but hurt lightly blurred ones (Guo et al., 2019). This directly contradicts the simplistic claim that “less blur is always better” for tracking.

For deblurring itself, blur shift appears as mismatch between training blur distributions and deployment blur distributions. ID-Blau addresses this with an implicit diffusion-based reblurring augmentation conditioned on a pixel-wise blur map $$I_{obj,b}(x) = \int K_x(\delta)\,I_0(x+\delta)\,d\delta.$$1 encoding orientation and magnitude. The method generates 10,000 additional blurred images and improves average performance across four backbones by +0.45 dB on GoPro, +0.48 dB on HIDE, +0.46 dB on RealBlur-J, and +0.50 dB on RealBlur-R (Wu et al., 2023). The underlying claim is not that blur can be perfectly parameterized by a single kernel, but that expanding the sampled blur-condition space reduces train–test mismatch.

6. Blur shift spaces in symbolic dynamics

Outside imaging, “blur shift” has a precise meaning in symbolic dynamics. “Blur shift spaces” introduces a compactification scheme for one-sided shift spaces over infinite alphabets by adding blur symbols, each representing an infinite subset of the alphabet. Sequences are identified once they agree up to the first blur symbol and then remain blurred forever by that symbol. The resulting space has a zero-dimensional topology generated by generalized cylinders of the form $$I_{obj,b}(x) = \int K_x(\delta)\,I_0(x+\delta)\,d\delta.$$2 and $$I_{obj,b}(x) = \int K_x(\delta)\,I_0(x+\delta)\,d\delta.$$3, is Hausdorff and regular, and generalizes both the Ott–Tomforde–Willis and Gonçalves–Royer constructions (Almeida et al., 2020).

This compactification changes the dynamical picture in a controlled way. The left shift is continuous away from pure blur points, but may be discontinuous on the level consisting entirely of blur symbols. Compactness, local compactness, first countability, second countability, separability, and metrizability admit explicit criteria in terms of the chosen blurred sets and the underlying language (Almeida et al., 2020). In this setting, “blur” has nothing to do with optics; it denotes deliberate topological coarse-graining of infinitely many symbols into one representative state.

“Maximizing measures for countable alphabet shifts via blur shift spaces” turns this construction into an ergodic-optimization tool. For a finite resolution $$I_{obj,b}(x) = \int K_x(\delta)\,I_0(x+\delta)\,d\delta.$$4, the paper defines blur-invariant probabilities for the measurable but generally discontinuous shift, proves the convex characterization

$$I_{obj,b}(x) = \int K_x(\delta)\,I_0(x+\delta)\,d\delta.$$5

and shows weak-* compactness under the finite cyclic predecessor assumption and denseness of periodic measures (Garibaldi et al., 24 Jul 2025). It then proves an existence theorem for maximizing probabilities for upper semi-continuous, bounded-above potentials $$I_{obj,b}(x) = \int K_x(\delta)\,I_0(x+\delta)\,d\delta.$$6 on $$I_{obj,b}(x) = \int K_x(\delta)\,I_0(x+\delta)\,d\delta.$$7 under the condition

$$I_{obj,b}(x) = \int K_x(\delta)\,I_0(x+\delta)\,d\delta.$$8

This is a conceptually different use of “blur shift,” but it preserves the same structural theme: the introduction of a coarse symbol records escape into an infinite family while restoring compactness.

7. Unifying perspective and recurrent misconceptions

Across these literatures, blur shift consistently marks the breakdown of a naive global-kernel description. In motion-blur geometry, the PSF centroid, orientation, and extent vary with local motion and reference time. In parallax, radiance-field, turbulence, and CT models, the effective blur depends on depth, pose, view, or detector position. In recognition and tracking, blur shift is not a PSF quantity at all, but a distribution shift induced by altered image statistics. In symbolic dynamics, it is a compactification device for non-compact shift spaces (Gast et al., 2016, Torres et al., 2023, Vasiljevic et al., 2016, Garibaldi et al., 24 Jul 2025).

Several misconceptions recur. One is that blur shift is always directly measurable from a single blurred image; cepstral blind deconvolution shows that kernel centroid is not identifiable from magnitude-only information and must often be centered by convention [0609165]. Another is that operator ordering under turbulence is arbitrary; the cited Fourier-optics analysis identifies tilt-then-blur as the physically correct model (Chan, 2022). A third is that removing blur is uniformly beneficial for downstream tasks; tracking results show that light blur can regularize or suppress clutter, whereas heavy blur is harmful, and selective rather than unconditional deblurring is preferable (Guo et al., 2019).

Taken together, these works indicate that “blur shift” names a class of phenomena in which blur carries structured information rather than merely corruption. In some settings that structure is geometric, in others statistical, and in others topological. The common thread is that blur is informative precisely when its variability is modeled rather than averaged away.