A laser scanner or depth camera samples a surface at discrete points. Inspection research usually treats a defect narrower than the point spacing as invisible, as if it fell between the samples. That rule is wrong in a way that matters. Chen and Li observed that pavement cracks narrower than their 4 cm terrain grid were sometimes still detected when they were long enough (A FEASIBILITY STUDY ON USE OF GENERIC MOBILE LASER SCANNING SYSTEM FOR DETECTING ASPHALT PAVEMENT CRACKS). Particle physicists would expect this. A weak signal spread across many measurements becomes detectable once enough measurements add up, provided the threshold is set honestly for how many places were searched.
This note proposes treating sub-resolution defect detection as a counting experiment and deriving an ideal-observer limit: the best detection rate any algorithm could reach at a fixed false-alarm rate, given the point spacing, the defect's width, length and depth, the sensor noise, and the size of the searched area. The limit makes two testable predictions that a pure geometric rule cannot. First, detectability depends on defect area rather than width alone. Second, converting raw points into a regular grid costs a predictable factor of sensitivity. The bound then serves as a yardstick. Running current learned 3D anomaly detectors against it under controlled downsampling tells us whether their failures on tiny defects are physical, so a better sensor is needed, or algorithmic, so a better model would help. Today that question is answered by intuition.
Core Claim
Take a linear defect of width $w$, length $L$ and depth $h$, sampled by points with mean spacing $d$ and independent height noise $\sigma$. About $wL/d^2$ points land inside the defect, each displaced by $h$. The optimal test for a known defect shape is a matched sum over those points, and its signal-to-noise ratio is
$$\mathrm{SNR}_{\text{points}} = \frac{h}{\sigma}\,\frac{\sqrt{wL}}{d}.$$
This follows directly from adding $wL/d^2$ independent Gaussian measurements. If the points are first averaged into grid cells of size $c > w$, as nearly every pipeline does with terrain models or voxels, each cell's signal is diluted by $w/c$ while its noise falls only through averaging. The best combination of cell means then gives
$$\mathrm{SNR}_{\text{grid}} = \mathrm{SNR}_{\text{points}}\sqrt{w/c}.$$
So binning costs a factor $\sqrt{w/c}$. This is the spatial version of the loss Ahmed et al. avoided by fitting unbinned likelihoods to continuous kinematic variables (Update of the search for the neutrinoless decay τ→μγ).
The defect's position and orientation are unknown, so the detector takes the maximum over roughly $M$ effectively independent placements. To keep the per-part false-alarm rate fixed, the required SNR therefore rises roughly as $\sqrt{2\ln M}$, the usual behaviour of the maximum of many Gaussian statistics. This is the look-elsewhere penalty that cut the Telescope Array cosmic-ray hotspot from $5.1\sigma$ to $3.4\sigma$ (INDICATIONS OF INTERMEDIATE-SCALE ANISOTROPY OF COSMIC RAYS...). Defect-detection papers apply the same scan over millions of patches and never account for it.
Together these give a detectability boundary in the width–length plane. On raw points the boundary is a hyperbola, $wL$ constant. On a grid it becomes $w^2L$ constant. In both cases it shifts outward logarithmically with the area searched. The proposal is to derive this boundary, extend it to realistic correlated noise, check it on real 3D industrial data, and use it to measure how much of the achievable sensitivity current detectors actually reach.
Relation to Prior Work
Three papers motivate the gap by documenting resolution failure without a statistical account. Chen and Li report the length-compensates-width effect on 3.4 cm mobile laser scans but explain it only informally. Huang et al. argue that crevasse extraction fails unless crevasse width substantially exceeds point spacing (A Bidirectional Analysis Method for Extracting Glacier Crevasses from Airborne LiDAR Point Clouds). Zhu et al. state that downsampling their 0.002 mm 3D scans makes detection of pits and scratches "plummet", but give no number (Real-IAD D3). All three treat the limit as a geometric cutoff, whereas we model it as a hypothesis test whose outcome depends on accumulated evidence.
Two physics papers supply the method. The unbinned extended maximum-likelihood fit of Ahmed et al. motivates testing directly on point coordinates. The trial-factor calibration of Abbasi et al. motivates correcting thresholds for the number of positions searched.
The EXO-200 analysis comes closest to joining the two directions (Search for Neutrinoless Double-β Decay with the Complete EXO-200 Dataset). There a neural network reads event topology from images on a wire grid and raises signal efficiency to 96.4%. It treats wire pitch as fixed and never asks how performance degrades as the grid coarsens. That degradation is exactly the curve we propose to predict and measure.
Learned multimodal detectors constrain the design. PIRN addresses scarcity of training samples (PIRN), and M3DM-NR addresses contaminated training sets (M3DM-NR). Both take sensor resolution as given and report only AUROC (area under the receiver-operating curve). MVTec AD shows how misleading that summary can be: a cable model scores 0.86 AUROC but only 0.05 per-region overlap (MVTec AD). We therefore report detection rate at a fixed per-part false-alarm rate, stratified by defect size relative to spacing.
Research Program
The first component is theory. We extend the two SNR expressions beyond white noise. Real scanner noise is spatially correlated, so we estimate its covariance from defect-free scans and whiten before matched filtering. A regular grid differs from a randomly jittered one: on a regular grid a defect narrower than the spacing can fall wholly between rows, so detection depends on its phase relative to the grid. We handle this by averaging over phase, which predicts a steeper and more variable transition on regular grids. The search count $M$ is estimated by Monte Carlo on normal parts, in the way Telescope Array calibrated its trial factor.
A Gaussian-noise observer would be too optimistic on real parts, because benign structure such as edges, seams and texture is not white noise. The NHA12D study documents this failure on highways, where DeepCrack reached 0.966 recall but only 0.141 precision because expansion joints look like cracks (NHA12D). We therefore define a second, background-limited observer. It uses a patch-level background model fitted to normal scans and is the stronger yardstick. The gap between the two observers measures how much of the problem comes from structured mimics rather than sampling.
The second component is a controlled testbed. Real-IAD D3 provides 3D scans at 0.002 mm point precision across 20 categories, with 5,000 normal and 3,450 anomalous samples. On held-out normal scans we inject synthetic grooves and pits with known $w$, $L$ and $h$. We then resample each scan to spacing $d$ under two schemes, random thinning and regular-grid resampling, over a range from native resolution to several times the smallest defect width. This gives ground-truth detectability surfaces over $(w/d, L/d, h/\sigma)$. Real annotated defects provide a second check: we measure their mask extents, place them in the same coordinates, and test whether the predicted boundary separates detected from missed real defects.
The third component is measuring detectors against the bound. For each detector and each spacing we fit the minimum SNR it needs for 50% detection at a fixed per-part false-alarm rate. We define its sampling efficiency as the squared ratio of the ideal observer's required SNR to the detector's required SNR. An efficiency near one means only better sensors can help. A low efficiency that falls as spacing coarsens means the model, not the physics, is the bottleneck.
Baselines
The naive reference is the geometric rule implicit in the inspection literature: a defect is declared detectable if and only if $w \ge d$. We score it as a predictor of which injected and real defects each method detects. The bound must predict detection outcomes more accurately than this rule, measured by balanced accuracy over held-out defects. It must also correctly predict detection of narrow-but-long defects, the case Chen and Li observed and the rule gets wrong.
The classical signal-processing baseline is the pipeline of Chen and Li: grid the points into a terrain model, high-pass filter, then apply 12 oriented Gaussian matched filters, with the threshold set at the same per-part false-alarm rate on the same normal scans. This is the strongest hand-built competitor, because it already is a matched filter and differs from our observer only in gridding and threshold calibration. Its required SNR should exceed the raw-point observer's by roughly $\sqrt{c/w}$ for sub-cell defects. That ratio is our test of the binning prediction.
The strongest learned method is D³M on Real-IAD D3. Zhu et al. report image-level AUROC 0.890 and pixel-level AUROC 0.937 at native resolution. For downsampled inputs, no reported result was found on Lacuna. We also run PIRN, which reports +3.9 image-level AUROC points over the strongest prior baseline in the 5-shot MVTec 3D-AD setting. We run M3DM-NR as well, which reports 86.6% image-level AUROC in the 10%-contamination MVTec 3D-AD setting. For a fair comparison, every learned model is retrained at each spacing on identically resampled normal data, using its published hyperparameters and an equal tuning budget of the same number of configurations per spacing. Every model's threshold is calibrated on the same held-out normal parts.
On this question, "beating" a learned detector does not mean building a better detector. It means the bound predicts each detector's failure spacing and failure shape, $wL$ versus $w^2L$, more accurately than the geometric rule. A secondary success would be showing that a simple raw-point likelihood test equals or exceeds a learned detector's detection rate in the sub-resolution regime.
Two ablations isolate the key ingredients. One compares the ideal observer on raw points against the same observer on gridded data, which isolates the binning penalty. The other compares a search-corrected threshold against a per-patch threshold, which isolates the look-elsewhere correction.
Why not just…?
Why not just sweep the downsampling factor and plot AUROC against resolution? This is the most serious alternative, and we run it. An empirical sweep says where a given detector fails. It cannot say whether a better algorithm could do better at that spacing, and it cannot extrapolate to defect shapes or search areas outside the sweep. The bound adds value only if it predicts. The decisive test fits the boundary on one set of categories and defect shapes, then checks predicted 50%-detection contours on held-out categories and on elongated defects. If curves interpolated from the sweep predict held-out outcomes as well as the bound does, the theoretical part of the proposal has failed, and we will report that.
Why not just use a higher-resolution sensor? Often one cannot. Post-disaster triage needs wide coverage, and Bai et al. show accuracy on low-resolution field images capped at 67.6% against 81.1% on high-resolution images (DETECTING CRACKS AND SPALLING AUTOMATICALLY IN EXTREME EVENTS...). In other settings the physics fixes the resolution: ultradeep seismic surveys (Application of a dynamic optimization-based multi-attribute fusion method for fault detection) and orbital radar (DETECTION OF BUILDING HEIGHT AND COLLAPSE MODE...). Even where a finer sensor is available, the bound answers the purchasing question of how much finer spacing is needed to reach a target miss rate for a given defect class.
Why not just set thresholds empirically on normal parts instead of deriving a look-elsewhere correction? For thresholding, this works, and we adopt it. All operating points in the proposal are calibrated on held-out normal scans. The analytic $\sqrt{2\ln M}$ term is kept only to predict how the required SNR grows when the inspected area grows beyond what normal data can calibrate. We test that prediction by calibrating on small crops and checking false-alarm rates on full parts.
Why not trust deep backbones to learn multi-scale detection end to end? They may well approach the bound on some categories, and that would be a useful finding in its own right. The evidence that they need not is the precision collapse on NHA12D and the localization collapse on MVTec AD. The efficiency measurement settles the question per category instead of leaving it to assumption.
Decisive Checks
The proposal succeeds if three things hold on Real-IAD D3 with injected defects. First, the measured 50%-detection contours for the raw-point observer follow $wL$ constant, and those for the gridded observer follow $w^2L$ constant, within fitted error. Second, the bound predicts detection of held-out real defects more accurately than the $w \ge d$ rule, including narrow-but-long defects. Third, the efficiencies of D³M, PIRN and M3DM-NR can be estimated tightly enough to separate categories where these detectors are near the limit from those where they are far below it.
The proposal fails if contour shapes are explained no better by the theory than by smooth interpolation of the sweep. It also fails if structured background dominates so completely that the Gaussian observer is uninformative even after whitening. In that case the background-limited observer becomes the only useful yardstick, and the paper reports that sampling theory contributes little beyond it.
As a secondary, qualitative check, if the Chen and Li mobile-laser data or the Malaspina glacier scans are accessible, we will verify the area law on real sub-grid cracks. Huang et al. report crevasse F1 of 94.61% at native density, which gives the reference point for that test.
References
- A FEASIBILITY STUDY ON USE OF GENERIC MOBILE LASER SCANNING SYSTEM FOR DETECTING ASPHALT PAVEMENT CRACKS
- Update of the search for the neutrinoless decay<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mrow><mml:mover><mml:mrow><mml:mi>τ</mml:mi></mml:mrow><mml:mrow><mml:mo>→</mml:mo></mml:mrow></mml:mover></mml:mrow></mml:mrow><mml:mi>μ</mml:mi><mml:mi>γ</mml:mi></mml:math>
- INDICATIONS OF INTERMEDIATE-SCALE ANISOTROPY OF COSMIC RAYS WITH ENERGY GREATER THAN 57 EeV IN THE NORTHERN SKY MEASURED WITH THE SURFACE DETECTOR OF THE TELESCOPE ARRAY EXPERIMENT
- A Bidirectional Analysis Method for Extracting Glacier Crevasses from Airborne LiDAR Point Clouds
- Real-IAD D3: A Real-World 2D/Pseudo-3D/3D Dataset for Industrial Anomaly Detection
- Search for Neutrinoless Double-<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mi>β</mml:mi></mml:mrow></mml:math>Decay with the Complete EXO-200 Dataset
- PIRN: Prototypical-based Intra-modal Reconstruction with Normality Communication for Multi-modal Anomaly Detection.
- M3DM-NR: RGB-3D Noisy-Resistant Industrial Anomaly Detection via Multimodal Denoising
- MVTec AD - A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection
- NHA12D: A New Pavement Crack Dataset and a Comparison Study Of Crack Detection Algorithms
- DETECTING CRACKS AND SPALLING AUTOMATICALLY IN EXTREME EVENTS BY END-TO-END DEEP LEARNING FRAMEWORKS
- Application of a dynamic optimization-based multi-attribute fusion method for fault detection
- DETECTION OF BUILDING HEIGHT AND COLLAPSE MODE FROM THE MEASUREMENT OF SHADOW LENGTH IN OPTICAL IMAGES AND LAYOVER IN SAR INTENSITY IMAGES