Most keypoint detectors find points to match between images by blurring each image at a series of increasing strengths, called a scale space, and looking for points that stand out at some blur level. KAZE showed that edge-aware blurring helps. KAZE replaces the Gaussian blur with nonlinear diffusion, a smoothing process that slows down wherever the image gradient is large, and this gains repeatability under noise. KAZE's smoothing still senses only how strong an edge is, not how sharply it bends. As a result it rounds corners, straightens tight curves and merges crossing lines at coarse scales. Those are the structures a corner-oriented detector most needs to keep.
We propose one change. The KAZE conductance, meaning the local smoothing rate, is multiplied by a second factor that falls toward zero wherever the local edge direction turns quickly. The direction is estimated from a smoothed structure tensor, the local average of gradient outer products, rather than from raw second derivatives. This gives a scale space in which corners and junctions stay in place across scales while straight edges and flat regions are smoothed as before. Because the conductance stays a scalar, KAZE's fast solver still applies, and the cost should stay close to KAZE's. Because the curvature sensor uses doubled orientation angles, it does not change when contrast flips, which is the failure mode of infrared-to-visible matching.
Core Claim
A nonlinear scale space whose conductance is gated by a regularized, polarity-invariant curvature signal will localize corners and junctions more stably across scale than first-order nonlinear diffusion. We expect this to improve detector repeatability and matching recall over KAZE under noise and contrast inversion at comparable cost. We expect no gain, and possibly a loss, on blob-dominated and texture-dominated scenes. The proposal stands or falls on whether the gain is specific to corner-rich, noisy or cross-modal content. If the gain cannot be separated from simply tuning KAZE's contrast parameter, the proposal fails.
The mechanism comes from curvature-adaptive diffusion. Feature-Preserving Flows: A Stochastic Differential Equation's View treats level-line smoothing as a particle random walk along edges. It stops the walk where the normal angle changes fast, $\|\nabla\theta\|$ large, and this preserves star tips and T-shapes that Perona–Malik rounds off. That work gives us the control signal. We do not reuse its random-walk discretization, whose bilinear interpolation the authors' setting shows slowly reintroduces rounding. Curvature and Singularity Driven Diffusion for Oriented Pattern Enhancement with Singular Points uses an exponential curvature penalty on along-flow diffusivity to keep fingerprint cores intact. We borrow its functional form. Neither paper builds a scale space, detects points or matches anything; their outputs are cleaned images.
KAZE Features supplies the detection machinery: additive operator splitting for large stable time steps, and scale-normalized Hessian extrema. It also supplies the first-order conductance $g(\|\nabla L_\sigma\|)$ that we argue is incomplete.
Research Program
Let $L(x,y,t)$ be the image at diffusion time $t$. Let $J_\rho$ be its structure tensor smoothed at integration scale $\rho$, and let $\phi$ be the doubled-angle dominant orientation of $J_\rho$. The doubled angle maps a gradient and its 180°-flipped twin to the same value. The evolution is
$$\frac{\partial L}{\partial t} = \mathrm{div}\!\left( g\!\left(\|\nabla L_\sigma\|\right)\, h\!\left(\tfrac{1}{2}\|\nabla \phi\|\right)\nabla L \right),\qquad h(s)=\exp\!\left(-s^2/k_\kappa^2\right).$$
Here $g$ is KAZE's $g_3$ conductance, $\tfrac{1}{2}\|\nabla\phi\|$ estimates the turning rate of edge direction, and $k_\kappa$ is a curvature threshold. We will set $k_\kappa$ adaptively as a percentile of the turning-rate histogram, the same way KAZE sets its contrast threshold. The gradient $\nabla\phi$ is computed through the unit vector $(\cos\phi,\sin\phi)$ to avoid wrap-around. It is also weighted by structure-tensor coherence, so that isotropic noisy regions do not register as high curvature.
This design answers the main known weakness of curvature flows: second derivatives amplify noise quadratically. Our curvature signal never differentiates the image twice directly. It differentiates an orientation field that has already been averaged over a window.
The build has three parts. The first is the gated scale space inside an open KAZE-style pipeline, keeping its octave layout, solver, Hessian detector and M-SURF descriptor fixed so that only the conductance changes. The second is a diagnostic suite on synthetic shapes: stars, T-, L- and X-junctions, and tight arcs under graded Gaussian noise. On this suite we measure two quantities against ground-truth vertices: how far each detected corner drifts as scale increases, and whether junctions survive to coarse octaves. The third is a hybrid detector variant. It pools Hessian-determinant extrema with Harris-type responses computed on the gated scale space instead of on Gaussian-smoothed images. This tests whether the decoupled corner detectors of Bai et al. and Hou et al. benefit from a better front end.
There is a risk we will measure rather than assume. Freezing smoothing at corners keeps them sharp at every scale, which may flatten the scale-normalized Hessian response and weaken scale selection. We will therefore report position repeatability and scale-estimate agreement separately.
Baselines
The naive reference is the linear Gaussian scale space with the same Hessian detector and M-SURF descriptor, plus standard SIFT. It shows how much of any gain is due to nonlinear diffusion at all.
The strongest handcrafted competitor is KAZE with its $g_3$ conductance. It is the closest existing method and has reported gains over classical detectors. On the Iguazu noise set, KAZE Features reports repeatability "up to 20% higher than SURF/STAR and 40% higher than SIFT." On the Mikolajczyk sequences it reports recall "40% higher than SURF, SIFT, and STAR for the same number of detected keypoints." These are relative figures, so we will rerun KAZE ourselves. For fairness, both methods use identical octaves, time steps, detector thresholds adjusted to equal keypoint counts, and the same descriptor. Both get the same hyperparameter search budget: KAZE's contrast percentile alone, and for our method the contrast and curvature percentiles together, searched over the same number of trials and selected on held-out sequences. Beating KAZE means higher repeatability and higher matching recall at equal keypoint count on the noise-heavy and corner-rich sequences, with confidence intervals across sequences that exclude zero, and no significant loss elsewhere.
The ablations isolate the key ingredient. The first gates with raw image curvature computed from second derivatives instead of the structure-tensor orientation, to test the noise-robustness argument. The second uses the curvature gate alone without the gradient factor $g$. The third uses single-angle instead of doubled-angle orientation, to test the polarity argument on cross-modal pairs.
For context we report LoFTR. LoFTR reports a homography AUC@3px of 65.9% on HPatches against 53.9% for SuperGlue. We do not expect a handcrafted detector to reach that level and do not claim it. No reported result was found on Lacuna for KAZE homography AUC on HPatches, so we will measure it.
For cross-modal pairs the reference is the HOFT pipeline of Zhu, Mao et al.. It reports an average of 9.78 seconds per infrared–visible pair. No raw correct-match counts or RMSE values were found on Lacuna, so we will reimplement HOFT on the same pairs.
Why not just…?
Why not just tune KAZE's contrast parameter? Lowering the contrast threshold makes KAZE stop smoothing at more edges, which also slows corner rounding. This is the most serious threat, because the gradient factor and the curvature factor are correlated at corners. We settle it directly. KAZE gets a full sweep of its threshold, and we plot repeatability against keypoint count for every setting. The curvature gate counts as a contribution only if it moves results beyond the envelope KAZE traces. If it does not, we will say so. The physical argument for a gain is that straight high-contrast edges and high-contrast corners have the same gradient magnitude, so no single threshold treats them differently, whereas turning rate does.
Why not just denoise with a curvature-preserving flow first and then run SIFT? A two-stage pipeline would apply the flow of Unal et al. once and then build a Gaussian pyramid on top. The linear pyramid would still move corners at coarse octaves, the drift that Image Visualization and Restoration by Curvature Motions attributes to linear filtering. We include this pre-filter baseline with the same flow parameters. The scale-drift diagnostic on synthetic vertices shows whether building the gate into every octave matters or whether one-shot denoising is enough.
Why not just use a learned matcher? On standard benchmarks, LoFTR's reported numbers dominate handcrafted pipelines. We do not dispute that. Our target settings are those where learned dense matchers are unavailable or untrained: embedded front ends like those in mine-tunnel stitching, and cross-spectral pairs like the power-inspection data, where there is little training data. Guiding Local Feature Matching with Surface Curvature adds curvature to learned matchers after detection, using 3D quadrics fitted to monocular depth, and reports AUC@5° gains of 1.0–1.5% on ScanNet. That approach operates after matching and depends on depth estimates. Ours works in the 2D image before detection and needs no depth network. The two are complementary, not competing.
Why not just lift to orientation scores? The crossing-preserving diffusion of Franken and Duits handles X-junctions better than any scalar flow, but its cost grows with the number of orientation slices. On the synthetic X-junction suite we run it as an accuracy ceiling. We expect it to win on crossings and lose badly on runtime. The proposal survives only if our scalar gate recovers a meaningful share of that accuracy at close to KAZE's cost.
Proposed Experiments
The first experiment is the synthetic diagnostic. We measure corner drift across octaves and junction survival under noise. The prediction is lower drift than KAZE and the Gaussian baseline, and a growing advantage as noise increases. If raw-curvature gating beats structure-tensor gating here, the noise argument is wrong.
The second experiment is detector repeatability and matching recall at fixed keypoint counts. It uses the Iguazu noise set and the Mikolajczyk sequences (Bark, Boat, Graf, Leuven, Trees, UBC, Wall), with homography AUC and mean matching accuracy on the HPatches illumination and viewpoint splits. The prediction is gains on the corner-rich sequences and on noise, and parity on texture-heavy Trees and Bark. Uniform gains everywhere would suggest the gate is acting as a hidden contrast retune, and we would check that against the KAZE sweep.
The third experiment is cross-modal. It uses the twelve infrared–visible power-scene groups of Zhu, Mao et al., if obtainable, and otherwise synthetic contrast-inverted HPatches pairs, scoring correct-match count and RMSE after RANSAC. The doubled-angle ablation here is the decisive check for the polarity claim. Its intent parallels the squared-gradient averaging in PIIFD, which we credit as the descriptor-level precedent for orientation invariance under modality inversion.
Runtime per image is reported in every experiment. A slowdown of more than about 1.5 times KAZE would weaken the practical case even if accuracy improves.
References
- Feature-Preserving Flows: A Stochastic Differential Equation's View
- Curvature and singularity driven diffusion for oriented pattern enhancement with singular points
- KAZE Features
- Real-Time Video Stitching for Mine Surveillance Using a Hybrid Image Registration Method
- Satellite-Borne Optical Remote Sensing Image Registration Based on Point Features
- LoFTR: Detector-Free Local Feature Matching With Transformers
- An Infrared and Visible Image Alignment Method Based on Gradient Distribution Properties and Scale-Invariant Features in Electric Power Scenes
- Image Visualization and Restoration by Curvature Motions
- Guiding Local Feature Matching with Surface Curvature
- Crossing-Preserving Coherence-Enhancing Diffusion on Invertible Orientation Scores
- Robust point matching method for multimodal retinal image registration