Forgive the Shift, Not the Rounded Corner: A Curvature-Adapted Image Distance in Orientation Space

Image quality metrics fail people in two opposite ways. They punish harmless changes: a sharp edge moved by one pixel looks identical to a human, yet pixel error counts it twice, and deep metrics can flip their rankings over it. They also forgive harmful changes: a rounded-off corner, a straightened bend, or two crossing lines merged into a blob can score well, because the error is small in energy and texture statistics barely move. The usual fix for the first problem is low-pass filtering before comparison, through Gaussian pre-smoothing or anti-aliased pooling. That fix makes the second problem worse, because blurring is exactly the operation that rounds corners and merges crossings.

This note proposes a full-reference image distance that separates the two kinds of change instead of trading one for the other. Each image is first "lifted" into an orientation score, a stack of oriented filter responses indexed by position and edge angle. In that representation a small translation moves energy in position but leaves the angle coordinate untouched. Corner rounding, bend straightening and crossing merges all redistribute energy across angles. The distance therefore smooths generously in position and strictly limits smoothing in angle. The positional smoothing follows the curvature of the reference image's contours, so tolerance runs along curved edges rather than cutting across them. The proposal has two parts. The first is a human-judged diagnostic test of whether people actually penalize geometry change more than energy-matched displacement, which nobody has measured. The second is the metric itself, built from tools that curvature-adaptive diffusion developed for denoising and never applied to perceptual evaluation.

Core Claim

Isotropic smoothing makes a metric tolerant to shifts, but it cannot tell displacement from deformation, because both are high-frequency changes in the plane. Perceptual Fidelity Aware Mean Squared Error (PAMSE) shows that mean squared error computed on Gaussian-smoothed images lifts Spearman correlation on TID2008 from 0.7718 to 0.9162. Its filter is linear and isotropic, though, so a rounded corner and a shifted corner of equal energy look the same to it. The same holds for the shift-tolerant deep metric of Shift-Tolerant Perceptual Similarity Metric, which gets its stability from fixed isotropic blur-pooling inside the network.

Curvature-adaptive diffusion supplies the missing separation. Nonlinear Diffusion on the 2D Euclidean Motion Group lifts images to orientation scores and steers smoothing along circular arcs whose curvature is read from the score's Hessian. Total Variation and Mean Curvature PDEs on the Homogeneous Space of Positions and Orientations builds locally adaptive "gauge frames" (local coordinate axes aligned with the structure at each point) from the same Hessian. In both cases, crossings and tight bends survive because they are disentangled along the angle axis. Feature-Preserving Flows: A Stochastic Differential Equation's View shows that switching smoothing off where the edge normal turns quickly preserves vertices that unmodulated flows round into circles. These papers use the geometry to decide where to smooth an image. We use it to decide where a metric should forgive.

Write $U_f(x,\theta)$ for the orientation score of image $f$: the response at position $x$ of an invertible oriented wavelet at angle $\theta$, as in the motion-group papers above. For a reference $r$ and test image $g$, the distance is

$$D_r(g) = \sum_{\theta} \int \big| (S_r [\,U_r - U_g\,])(x,\theta) \big|^2 \, dx .$$

Here $S_r$ is a linear smoothing operator whose shape is fixed by the reference alone. At each point $(x,\theta)$ it averages over a spatial width $\sigma_s$ along the curvature-$\kappa_r(x,\theta)$ arc of the reference's gauge frame, and over a narrower width across it. In angle, it averages over a small width $\sigma_\theta$, multiplied by a gate that falls to zero where the reference has several distinct orientation peaks at one location, such as corners and junctions.

Because $S_r$ depends only on $r$, $D_r$ remains a quadratic, convex and differentiable function of $g$, just as PAMSE is. What is new is a reference-adapted quadratic form that is loose in position and strict in angle, with its looseness bent to follow the reference's contours.

Research Program

The first component is a diagnostic study, because the whole proposal rests on an untested premise. The curvature-diffusion papers prove that standard smoothing straightens bends and rounds corners. None of them tested whether observers find this more objectionable than displacement or noise of equal energy. We will build a set of distorted images from line-rich references: TID2013 and PIPAL reference images, photographs of buildings and text, and fiber microscopy crops of the kind used by Franken et al.

The distortions come in two families. Geometry-altering distortions are generated with the very flows the literature identifies as harmful: Gaussian or plain geometric-heat-equation smoothing localized at corners, Perona–Malik smoothing on tight bends, plain 2D coherence-enhancing diffusion at crossings, and radial warps. Geometry-preserving distortions are sub-pixel global shifts, small displacements along the contour, and mild noise. Every geometry-altering image is paired with a geometry-preserving one matched in both MSE and PAMSE. Human two-alternative forced-choice judgments are then collected on these pairs, following the protocol of BAPPS-style patch comparisons.

The second component is the metric. The steps are lifting with a fixed number of orientations (we will try 8, 12 and 16), gauge-frame and curvature estimation on the reference score using the Hessian construction of the motion-group papers, the junction gate, and the comparison. The operator $S_r$ is applied as separable convolutions along precomputed per-point arcs. We do not run a time-stepped flow, which sidesteps the small-time-step stability problem noted for explicit schemes on the motion group. The free parameters are $\sigma_s$, $\sigma_\theta$ and the gate threshold, set once globally on a held-out split. Runtime will be reported as a multiple of MSE, against PAMSE's reported 2.6 times.

The third component is a check on real outputs. We will evaluate restorations of noisy line-rich images produced by classical and curvature-adaptive methods and ask whether the metric's rankings match human rankings where PSNR and LPIPS do not.

Baselines

The naive reference points are MSE and PSNR. They show how large the gap is before any correction: PAMSE's paper reports MSE at 0.7718 Spearman correlation on TID2008, and PIPAL reports PSNR as anti-correlated with humans on GAN-restored images.

The most direct competitor is PAMSE, because our distance is the same kind of object, a smoothed quadratic error, with a different smoothing operator. If anisotropic, curvature-adapted smoothing adds nothing, PAMSE will tie. It will be run on identical images, with its Gaussian width tuned on the same held-out split and the same number of tuning trials as our three parameters. On TID2008 the bar is its reported 0.9162. Beating it means staying within about 0.01 there while winning clearly on the geometry-altering pairs and on TID2013's geometric distortion types.

The strongest deep competitor on shift robustness is the anti-aliased shift-tolerant LPIPS of Shift-Tolerant Perceptual Similarity Metric. It reports a 1-pixel rank-flip rate on BAPPS of 0.57%, against 2.81% for original LPIPS, and 76.97% accuracy on CLIC against 73.68%. On PIPAL's GAN-based distortions, PIPAL reports LPIPS-Alex at 0.4862 correlation, rising to 0.5135 with blur-pooling and no retraining. We will run this metric with its published weights. Beating it means a flip rate no worse than 0.57% on Shifted BAPPS together with higher agreement on the geometry pairs. We will report PIPAL, but we do not expect an analytic metric to beat trained deep metrics on GAN texture artifacts. Being competitive there is enough.

E-LPIPS is the strongest robustness-by-randomization method, reporting 69.16% two-alternative accuracy on BAPPS. It will be run at its recommended number of random samples, and runtimes will be reported side by side.

BiRQA is the strongest trained full-reference metric in the file, with Spearman correlation of 0.988 on LIVE and 0.979 on CSIQ. Its authors also admit weakness on TID2013's radial transforms, for which the file has no reported number. We include it if its weights are available. We do not claim to beat it on LIVE or CSIQ. The comparison that matters is the geometric subsets.

Three ablations isolate the key ingredient. The first lifts the images but uses straight, curvature-zero gauge frames, which tests whether curvature adaptation matters beyond plain lifting. The second skips lifting and applies curvature-gated variable-width smoothing in the plane, which tests whether the angle axis is what separates displacement from deformation. The third replaces our operator with Perona–Malik or bilateral pre-filtering, which tests whether any edge-aware smoothing would do.

Why not just…?

Why not just tune or stack PAMSE's Gaussian at several scales? Any linear isotropic filter in the plane weights an error by its spatial frequency alone. A sub-pixel corner shift and a corner rounding of equal energy occupy overlapping frequencies, so no choice of width can rank them apart. The diagnostic set makes this concrete: its pairs are matched in PAMSE by construction, so a multi-scale PAMSE can only agree with humans on them by chance. If humans also fail to separate the pairs, the premise is wrong, and the proposal should stop there.

Why not just use shift-tolerant LPIPS or E-LPIPS, whose networks already contain oriented filters? This is a real threat. Early convolutional layers do respond to orientation, and these metrics may already penalize corner rounding. Against this, Xiaoshi Wu and the PIPAL analysis report that deep feature metrics are dominated by texture and miss structural errors, and the shift-tolerant design adds isotropic blur that should spread angle-specific evidence. The geometry pairs settle the question directly. If shift-tolerant LPIPS matches human choices there as well as our distance does, the contribution shrinks to an analytic, convex and interpretable alternative, and we will report it that way.

Why not just fine-tune a learned metric on geometric distortions, or add Hessian features to a BiRQA-style network? This is also a real threat once human labels exist. We will fine-tune shift-tolerant LPIPS on part of the diagnostic set and evaluate on held-out distortion families. For example, it trains on corner rounding and is tested on crossing merges and radial warps. We will also check that it keeps its BAPPS and PIPAL scores. If it generalizes across held-out families, a learned route is enough. Our distance's advantage would then lie only in needing no geometry-specific training data and in being usable as a convex loss.

Why not just use classical edge-preserving smoothing such as Perona–Malik before comparing? The motion-group and stochastic-flow papers show that first-order edge-stopping treats tight bends and corners as noise and produces checkerboard artifacts at crossings. A metric built on it would inherit the same blind spot. This is the third ablation above, and we expect it to fail specifically on crossing pairs.

Proposed Experiments

The decisive first experiment is the human study. If observers do not prefer geometry-preserving distortions over geometry-altering ones at matched MSE and PAMSE, the motivating gap is not real. Our hypothesis is a clear preference, concentrated at corners and crossings. Alongside it we will measure how often each baseline metric agrees with these human choices.

The second experiment scores all metrics on TID2008, TID2013 (overall and per distortion type, with emphasis on geometric types), LIVE, CSIQ, PIPAL, BAPPS two-alternative accuracy, and the Shifted BAPPS flip rate. Success means four things together: TID2008 correlation within about 0.01 of PAMSE's 0.9162, higher correlation than every baseline on TID2013's geometric types and on the diagnostic pairs, a flip rate at or below 0.57%, and a runtime within a small multiple of PAMSE. The ablations show which ingredient carries any gain. If the straight-gauge variant matches the full method, curvature adaptation is unnecessary, and we will say so.

The third experiment ranks restorations of noisy line-rich images: Gaussian, Perona–Malik, the curvature-controlled flow of Feature-Preserving Flows, CED-OS, Gauge TVF and BM3D. Each output is ranked by humans and by every metric. The prior evidence here is only PSNR, namely Gauge TVF's 26.1 dB against BM3D's 23.9 dB on the noisy "Mona Lisa" image. It is unknown whether people agree with that ordering, or whether our distance tracks their ranking more closely than PSNR and LPIPS do.

As a final check that convexity is useful, a small denoising network will be trained with our distance as its loss, against PAMSE and LPIPS losses. Its outputs will be compared by humans on held-out line-rich images.