When an engineer inspects a concrete wall or a sewer pipe, the useful output is a number: how wide the worst crack is, how long the crack network runs, and whether either exceeds a repair limit. Current vision pipelines produce a pixel mask, which is a map labeling each pixel as crack or not crack. They then measure width and length on that mask as if it were exact. For thin cracks on textured surfaces the mask is far from exact. The crack is camouflaged, meaning it matches the aggregate, stains, or brushed metal around it, so its edges are faint. On top of that, the human labels used to train the model systematically leave out thin branches and tips. A one-pixel boundary error on a three-pixel-wide crack is a 30% error in width, and the pipeline reports none of it.
We propose to treat the engineering measurement, not the pixel, as the object of uncertainty. The system estimates crack width at each cross-section and length along the skeleton, together with an interval around each value. It spends its high-resolution refinement budget only on the few cross-sections and branch junctions whose refinement would most shrink the interval on the reported measurement. The intervals are then calibrated against measurement ground truth that does not come from human masks. The claim we will test is narrow: under camouflage and label bias, measurement-targeted refinement plus measurement-level calibration gives width and length intervals that cover the truth at their stated rate. Pixel-entropy refinement and mask-sampling baselines do not, at equal compute.
Core Claim
Prior work leaves three pieces disconnected.
The first piece is metric sizing. Binocular Video-Based Automatic Pixel-Level Crack Detection and Quantification Using Deep Convolutional Neural Networks for Concrete Structures measures length by summing pixels along the crack's centerline (its skeleton). It measures maximum width with a "median plumb line" drawn perpendicular to that centerline. It reports 4.5% relative width error and 0.6-pixel average absolute width error, but only on 20 clean images of isolated cracks. It carries no uncertainty, so segmentation failures pass straight into the numbers.
The second piece is uncertainty on crack boundaries. Mind marginal non-crack regions: Clustering-inspired representation learning for crack segmentation predicts pixel variance in the ambiguous zone next to cracks, but uses it only during training and discards it at inference. Distribution-aware Noisy-label Crack Segmentation shows that annotators under-label thin cracks. It counters this with a Gaussian-mixture model of crack and background features, but still outputs a binary mask.
The third piece is selective refinement. Boosting the Dual-Stream Architecture in Ultra-High Resolution Segmentation with Resolution-Biased Uncertainty Estimation sends only the Top-K most uncertain patches to a high-resolution stream. Gate to the Vessel: Residual Experts Restore What SAM Overlooks (FineSAM++) adds a gate to the Segment Anything Model (SAM). The gate routes the regions where SAM is confused, typically on thin vessels it tends to break, to small corrective "residual expert" networks. Both papers choose where to refine by pixel-level confusion and score the result by overlap or connectivity of the mask.
What is new here is choosing where to refine, and judging success, by the measurement the engineer reads off. A high-entropy pixel on a long, wide, unambiguous crack barely changes its maximum width. A faint pixel at the narrowest cross-section, or a missed junction that splits one crack into two, changes the reported numbers a lot. The same logic gives a principled stopping rule for camouflaged regions. Where refinement is predicted not to reduce measurement variance, the system widens the interval instead of spending compute. This addresses the open problem of separating irreducible ambiguity (camouflage, sensor noise) from reducible ambiguity (resolution, limited data).
Research Program
Base model. A crack segmenter, initially the CrackMamba design from Topology-aware mamba for crack segmentation in structures or a SAM-Adapter backbone, runs on a downsampled image. It outputs a per-pixel crack probability map. From the thresholded map we extract the skeleton and cut it into cross-sections. We split branched networks at junctions so that Y-, X- and mesh topologies, which the binocular pipeline cannot handle, are measured branch by branch.
Measurement uncertainty. A light head predicts, for each cross-section $j$, a width estimate $\hat w_j$ and a spread $\sigma_j$. A second head predicts $\Delta_j$, the expected reduction in $\sigma_j$ if the surrounding patch were re-segmented at full resolution. Both heads are trained on pairs of low- and full-resolution predictions of the same image, an idea borrowed from the "resolution bias" term in the dual-stream paper.
Routing. We define the measurement that matters as the maximum width $W=\max_j w_j$, plus total length. The refinement score for the patch around cross-section $j$ is
$$s_j = \Delta_j \cdot P\left(j = \arg\max_k w_k\right),$$
which multiplies the expected variance reduction by the probability that cross-section $j$ is the one that sets the maximum width. That probability is estimated from the $\hat w$ and $\sigma$ values of all cross-sections. Skeleton gaps whose closing would merge or split branches receive an analogous score for length.
Refinement and output. Under a fixed budget, the top-scoring patches go to a high-resolution refinement module in the style of FineSAM++'s residual experts. We then output each measurement with an interval $[\hat W - q\,\sigma_W,\ \hat W + q\,\sigma_W]$. Here $\sigma_W$ is the propagated spread of the maximum width, and $q$ is a scale factor. We set $q$ by conformal calibration, a distribution-free procedure that picks $q$ from residuals on a held-out set so that intervals cover the truth at a chosen rate, for example 90%. The calibration residuals are measured against independent width truth, not against human masks.
Measurement truth. Building this ground truth is the main data contribution, because no benchmark in our file provides calibrated crack widths. It has three sources:
- Synthetic composites. We render cracks of known geometry onto real crack-free patches from Crack500, CrackSeg5k and SewerCrack. A camouflage dial matches the crack's contrast and frequency content to the background, using the ring-shaped frequency bands of Frequency-Guided Spatial Adaptation for Camouflaged Object Detection as the matching statistic.
- Repeat annotation. Several annotators independently re-label a subset of CFD and CrackTree200, which lets us measure how far human width and length readings spread.
- A physical set. A small set of concrete specimens is photographed and measured with a crack-width gauge. This is the only source not derived from images, and we must collect it ourselves.
Baselines
Naive reference. We train the base segmenter on the standard human masks, threshold it, and apply the skeleton and median-plumb-line measurement of the binocular paper. This is exactly how quantification is done today, and it reports point values with no interval. Its width error on our truth sets is the floor any method must beat. The binocular paper reports 4.5% relative width error on its own clean images. No reported result was found on Lacuna for width error on camouflaged or crack-benchmark images.
Strongest segmentation competitors. We use two strong segmenters, each followed by the same measurement step:
- Distribution-aware SAM-Adapter with the Gaussian mixture. On Crack500 it reports F1 of 76.66% and IoU of 60.56% in Distribution-aware Noisy-label Crack Segmentation. It is the strongest competitor because it directly targets the label bias that erodes thin cracks.
- CrackMamba. It reports 81.75% mean IoU and 79.32% F1 on CrackSeg9K, and 70.47% mean IoU on SewerCrack, in Topology-aware mamba for crack segmentation in structures.
Both are retrained on our training splits with the same augmentations and the same hyperparameter-search budget as our model. Beating them means lower width and length error at equal or lower compute. A higher mask F1 alone does not count.
Strongest interval baseline. We train a five-member ensemble of the base segmenter. We compute width and length on every member's mask and give the interval as the members' spread. We also run a conformally calibrated version that applies the same conformal step to this spread. Beating it requires our method to reach 90% interval coverage with narrower intervals, under matched compute.
Routing ablations. These isolate the key ingredient. Pixel-entropy Top-K routing follows the dual-stream paper's selector, uses the same refinement module, and gets the same patch budget. FineSAM++-style gating uses uncertainty-weighted residual experts on every pixel. FineSAM++ reports DRIVE Dice of 0.8231, and the dual-stream paper reports a 517x speedup over a cascade refiner. Neither has a reported crack-measurement result on Lacuna. If entropy routing matches measurement-sensitive routing on width error at equal budget, the core claim fails.
Why not just…?
Why not just compute widths from an ensemble and calibrate the spread? This is the most serious threat, and we include it as a full baseline rather than argue past it. Its weakness is predictable but has not been shown. The ensemble members share the same biased labels, so they agree with one another while all missing the same thin tips. Their spread therefore understates error exactly where camouflage is strongest. Conformal calibration can repair overall coverage by inflating every interval, but it cannot move the center of an interval that is biased. The decisive comparison is interval width at matched 90% coverage, broken down by camouflage level. If the calibrated ensemble matches us there, the routing contribution reduces to an efficiency claim, and we will report it that way.
Why not just segment everything at full resolution? On images like CrackSeg5k's 7360x4912 originals this costs far more compute, which is the problem the dual-stream paper's sparse routing exists to solve. More importantly, full resolution does not fix label bias. A full-resolution model trained on masks that omit thin branches still omits them. We run full-resolution inference as an accuracy upper bound to measure how much of it routing recovers at a fraction of the cost.
Why not just fix the labels with a noisy-label method and measure afterward? The distribution-aware method does improve thin-crack recall. However, it outputs a hard mask with no statement of how wrong the resulting width might be, and an engineer comparing a width against a repair limit needs exactly that statement. We will stack it under our approach: the label-corrected segmenter serves as our base model. If adding routing and calibration on top brings no gain in coverage or error, that is a negative result worth publishing.
Decisive Experiments
Calibration under camouflage. On the synthetic composites, we sweep the camouflage dial from high to near-zero contrast. We plot interval coverage against the 90% target and mean interval width, for our method, the ensemble baselines, and point estimators. The prediction is that our coverage stays near 90% as camouflage increases while intervals widen. Baselines should either lose coverage or widen uniformly everywhere.
Equal-budget routing. At patch budgets of 5%, 10% and 25%, we compare measurement-sensitive routing with entropy routing on maximum-width error and on branch-count error in networked cracks.
Comparison with human inspectors. On the repeat-annotated subset and the physical gauge set, we ask two questions. First, do our intervals contain the gauge reading more often than the range spanned by human annotators does? Second, how often does each method misclassify whether a crack exceeds a chosen width limit?
Transfer. We calibrate on concrete and test on SewerCrack, checking whether coverage holds without recalibration. Every result is reported per camouflage level, not only as an average.
References
- Binocular Video-Based Automatic Pixel-Level Crack Detection and Quantification Using Deep Convolutional Neural Networks for Concrete Structures
- Mind marginal non-crack regions: Clustering-inspired representation learning for crack segmentation
- Distribution-aware Noisy-label Crack Segmentation
- Boosting the Dual-Stream Architecture in Ultra-High Resolution Segmentation with Resolution-Biased Uncertainty Estimation
- Gate to the Vessel: Residual Experts Restore What SAM Overlooks
- Topology-aware Mamba for Crack Segmentation in Structures
- Frequency-Guided Spatial Adaptation for Camouflaged Object Detection