PaperlayerView exact revision on arXiv

Source-linked AI summary

Cross-View and Cross-Time Tracking of Crack Evolution in Concrete

Pengru Deng, Jiapeng Yao, Su Wang, Xuhui He, Zhao Wang, Takanori Emaru, Takashi Matsumoto, Xinrun Li

arXiv:2610.33550v1eess.IVcs.CV

TL;DR

Crack-growth monitoring needs to preserve physical crack identity across changing viewpoints and distinguish new extension from existing crack regions, beyond static crack detection. The paper proposes a geometry-assisted, morphology-aware tracking framework and reports loose identity correspondence under relaxed overlap, alongside a pooled instance-level CLCE of 0.753. Early-crack segmentation and projection errors constrain the results.

  • Problem

    Crack masks or geometric reconstruction alone do not preserve crack identity across views or quantify temporal extension, although monitoring requires tracking the same physical crack and separating new growth.

    Source: §1 · pp. 2–3

  • Method

    The framework combines geometry-assisted alignment, morphology-aware instance formation and identity propagation, and pairwise growth attribution in a unified reference view.

    Source: §1 · p. 5, §5 · p. 45

  • Results

    Across eight evaluated scenes, pooled instance-level CLCE was 0.753, and acceptable instance correspondence was achieved under a relaxed overlap criterion.

    Source: §5 · pp. 45–46

  • Takeaways & Limitations

    The framework extends fixed-camera crack-growth detection to identity-aware monitoring using overlapping images from arbitrary viewpoints.

    Source: §5 · p. 46

    Early micro-crack mask gaps and residual projection drift can fragment instances, reduce unified-view overlap, and delay growth localization; field generalizability remains unverified.

    Source: §4.5 · p. 44

Abstract

from arXiv · show

Crack propagation provides richer information than static crack observations for evaluating structural conditions of concrete infrastructure and planning maintenance strategies. Unlike static crack detection, which can now be largely automated using computer vision, vision-based crack tracking is a coupled cross-view and cross-time problem: the same physical crack must retain its identity as camera viewpoints change, while newly developed growth must be distinguished from existing crack segments in a unified reference view. Existing segmentation and 3D reconstruction methods can recover crack masks or scene geometry, but they do not preserve crack identity across views or quantify temporal extension. To address this gap, this study formulates crack growth monitoring as a cross-view and cross-time crack-instance tracking task and proposes a geometry-assisted, morphology-aware framework. The framework first uses image-based geometric registration to align observations from different viewpoints into a shared structural reference, then applies morphology-aware instantiation to assign physically interpretable crack identities, and finally performs temporal mapping in the unified view to distinguish newly developed growth from existing crack segments. To enable evaluation, a temporally annotated dataset was established and used, covering three representative cracking scenarios: single dominant crack propagation, mixed-mode branching, and multiple-crack interaction, with pixel-level masks and cross-time instance labels. Experimental results show that the proposed framework can establish loose cross-view identity correspondence under changing viewpoints, while strict spatial overlap and growth localization remain limited by early micro-crack segmentability and projection drift, highlighting directions for future crack-monitoring research.

1. Introduction

Crack-growth monitoring must track physical crack identity and extension across time despite changing viewpoints, whereas existing alignment and segmentation work largely addresses separate components. This study combines geometric alignment, crack-instance representation and association, and growth measurement, with an evaluation protocol and temporally annotated dataset.

1. Introduction

  • Vision-based crack-growth monitoring must determine whether observations show the same physical crack and how much new extension has occurred, not merely detect cracks in individual images.The task couples geometric alignment, instance identification, instance association, and growth separation and quantification.Source: §1 · pp. 3–4
  • Prior cross-view alignment methods mainly address single-time spatial consistency, while temporal tracking commonly assumes a fixed viewpoint, predefined camera path, or structure-specific landmark.The cited related work identifies a need to recover correspondence from overlapping images when camera pose changes freely, without those constraints.Source: §1 · p. 4
  • The proposed framework maps successive observations into a shared reference view, organizes aligned masks into crack instances, associates identities, and separates inherited regions from new growth.This integrates geometric registration with crack representation, association, and growth measurement.Source: §1 · p. 5
  • The study formulates crack-instance tracking under changing viewpoints to associate identities across time and detect growth from overlapping images acquired with freely chosen imaging devices.The images may be acquired from arbitrary viewpoints.Source: §1 · pp. 5–6
  • A proposed evaluation protocol assesses identity consistency, growth localization, and propagated-length quantification across viewpoints and time.Source: §1 · p. 6
  • The temporally annotated evaluation dataset provides pixel-level masks and cross-time instance labels for representative crack patterns.It supports joint evaluation of identity preservation and growth attribution.Source: §1 · p. 6

2.1 Framework of the proposed method

The framework treats cracks as structural entities tracked across stages in a shared geometric reference, rather than as isolated image patterns. Its workflow comprises geometry estimation, segmentation, instance identification, and crack numbering and extension tracking.

2.1 Framework of the proposed method

  • The four workflow modules are geometric structure estimation, image crack segmentation, crack-instance identification, and crack numbering and extension tracking.Fig. 1 presents the overall workflow.Source: §2.1 · p. 6, Figure 1 · p. 6

2.2 Geometric structure estimation

The geometric module registers observations from changing viewpoints into a shared structural reference, where aligned masks support instance association and growth separation. The method motivates dense, geometry-aware mappings because surface geometry and occlusion can make a single global planar mapping inadequate.

2.2 Geometric structure estimation

  • A small registration error can substantially reduce overlap for narrow, fragmented cracks and may be mistaken for temporal growth.The geometric module maps observations into a shared structural reference before instance association and growth separation.Source: §2.2 · p. 7
  • A single global homography may not describe pose-induced displacement across cracks spanning depth variation, non-planar geometry, multiple faces, or occlusion.The registration front end is therefore motivated to provide dense local support around narrow crack segments while accounting for supporting-surface geometry.Source: §2.2 · pp. 7–8
  • MASt3R provides dense image descriptors and 3D point maps for spatially varying mappings, reducing reliance on isolated keypoints and a single global planar model.The implementation uses its point maps, valid-pixel masks, camera poses, and focal lengths; the pretrained backbone is not retrained or modified.Source: §2.2 · pp. 7–8
  • The registration front end is modular: downstream tracking consumes the resulting mapping and confidence information, allowing other pose-recovery methods to replace the backend.The text identifies homography, SfM-derived camera poses, or dense correspondence with local 3D support as possible mapping forms.Source: §2.2 · pp. 8–9

2.3 Image crack segmentation

Crack segmentation supplies the visible crack support used by instance identification and temporal tracking, but is not the study’s main contribution. The evaluation uses a prior DeepLabv3+ checkpoint without retraining or fine-tuning on the evaluated temporal scenes.

2.3 Image crack segmentation

  • Predicted crack masks from a prior DeepLabv3+ checkpoint feed instance identification and temporal tracking without retraining or fine-tuning on evaluated temporal-scene images.Weak early micro-crack segmentation can propagate into fragmented instances, reduced cross-view overlap, and delayed growth localization.Source: §2.3 · p. 9

2.4 Crack instance identification

Crack-instance identification addresses connected components that mix distinct crack systems and disconnected fragments that belong to one physical crack. It skeletonizes components, uses morphology and geometry to split or merge branches, and groups compatible fragments while preserving foreground pixels.

2.4 Crack instance identification

  • A morphology-aware split-and-merge strategy addresses crossings that combine different crack systems and interruptions that fragment one physical crack.No component-area filtering is applied, so small fragments are retained.Source: §2.4 · p. 9
  • Each 8-connected component is skeletonized and decomposed into primitive branches, whose endpoints and principal directions support continuity analysis.When fewer than two reliable endpoints are available, the branch’s principal-axis extremes serve as representatives; direction estimates require at least four pixels and an extent of eight pixels.Source: §2.4 · pp. 9–11
  • Fragment merging combines endpoint proximity, directional agreement, and normal offset, with stricter directional tolerance as endpoint gaps grow.When direction estimates are unreliable, more restrictive distance and endpoint-axis alignment limits are used.Source: §2.4 · pp. 12–13
  • Branches crossing at angles of at least 45° are constrained to remain in separate instances when both dominant-direction estimates are reliable.A constrained union-find groups merge-compatible fragments while rejecting unions that would combine hard-separated branches.Source: §2.4 · pp. 13–14

2.5 Crack numbering and crack-growth tracking

The tracking module couples identity propagation and growth attribution in a shared geometric reference. It projects and matches instances, propagates identifiers over time, then measures newly appeared skeleton support as crack extension.

2.5 Crack numbering and crack-growth tracking

  • Identity propagation and growth attribution are treated as one coupled process in a shared geometric reference, following projection and instance matching.The module then numbers cracks chronologically and separates inherited skeleton support from newly appeared extension.Source: §2.5 · p. 16

2.5.1 Pose-guided crack projection and instance matching

  • Dilated-mask overlap provides an image-space tolerance for matching projected previous instances to current candidates despite minor projection and segmentation errors.A correspondence is accepted using intersection area, current-instance coverage, or Dice thresholds; the listed thresholds are 1 pixel, 0.02, and 0.05, respectively.Source: §2.5.1 · pp. 16–17

2.5.2 Crack numbering strategy

  • Current instances inherit the identifier of the highest-scoring accepted previous match, with nearest-track and single-instance fallbacks for weak or missing overlap.An instance that remains unassociated after these rules receives a new track identifier.Source: §2.5.2 · p. 19

2.5.3 Pairwise crack-growth extraction and length quantification

  • After pairwise registration, crack growth is measured as newly appearing skeleton pixels outside the inherited-support band, and extension length is the length of that skeleton.Skeleton comparison avoids conflating longitudinal extension with mask-width changes from crack opening, illumination variation, or segmentation noise.Source: §2.5.3 · pp. 19–20

3. Experiment Setup

The experiments use staged-loading observations with deliberate viewpoint changes and a dataset spanning three crack scenarios. Evaluation separates crack identity correspondence, growth-length consistency, same-interval localization, and delayed detection.

3.1 Data acquisition from staged loading experiments

  • Staged loading supplies real crack extension and freely changing viewpoints, while avoiding the environmental and acquisition confounds of multi-year field observations.The setup requires overlapping views around the crack region but prescribes no fixed camera distance.Source: §3.1 · pp. 21–22

3.2 Dataset construction

  • The dataset covers single-crack propagation, mixed-mode branching, and multiple-crack interaction across 53 images from eight specimens and 45 adjacent-stage transitions.Pixel masks and scene-local instance labels preserve physical crack identities across viewpoints and loading stages; growth references use manually reviewed geometric mappings.Source: §3.2 · pp. 22–23, §3.1 · p. 22

3.3 Evaluation metrics

  • Evaluation pairs HOTA for identity correspondence with CLCE for post-association growth-length consistency, supplemented by Immediate Growth F1 and DGR@K diagnostics.The protocol evaluates identity, growth localization, and propagated-length quantification; DGR@K includes reference growth detected in its interval or within the following K intervals.Source: §3.3 · pp. 24–25, §3.3.2 · pp. 27, 29

3.3.1 Growth-length consistency

  • CLCE compares predicted and verified newly appeared skeleton lengths after crack-instance association, excluding transitions with reference growth length at most 1 px.A small constant prevents division by zero, and CLCE does not measure spatial localization of predicted growth.Source: §3.3.1 · pp. 25–26

3.3.2 Growth localization and delayed crack-growth detection

  • Immediate Growth F1 assesses same-interval, growth-length-weighted instance detection, while DGR@K measures reference-growth coverage in that interval or within the next K intervals.Delayed matching searches intervals in increasing order and uses one-to-one assignment; each prediction can be assigned only once.Source: §3.3.2 · pp. 27–29

3.3.3 Crack-instance identity correspondence

  • HOTA assesses crack-instance identity correspondence across views and times, with HOTA(0) serving as a loose measure at the first IoU threshold of 0.05.Mean HOTA assesses stricter projected overlap, which can fall with small lateral displacement from segmentation uncertainty or projection drift.Source: §3.3.3 · pp. 29–30

3.4 Implementation details

  • The implementation keeps the segmentation model fixed, uses pretrained MASt3R for pairwise geometry, and applies morphology-aware instance identification and association across all eight scenes.Growth extraction uses 16 px image-space tolerances for projection error and mask discontinuity, not physical crack width.Source: §3.4 · pp. 30–31

4. Experimental Results and Discussion

Across the tested scenes, the framework preserves crack identities under loose cross-view overlap criteria, but strict spatial overlap and growth measurements remain sensitive to crack visibility and projection stability. Qualitative examples show identity continuity across changing views and crack-evolution scenarios, while delayed detection and ablation results further characterize its tracking behavior.

4.1 Qualitative analysis of crack-instance tracking

  • The pre-notched, L-shaped, and deep-beam examples preserve major crack identities across stages while distinguishing inherited segments from newly extended or branching segments.The examples cover stable-path lengthening, parent-crack branching, and interacting shear cracks under changing viewpoints and loading stages.Source: §4.1 · pp. 32, 34, Figure 7 · p. 33, Figure 8 · p. 34

4.2 HOTA-based identity-correspondence analysis

  • HOTA(0) ranges from 0.111 to 1.000 across Scenes 1–8, while stricter overlap thresholds rapidly reduce HOTA, DetA, and AssA in most scenes.The results support loose cross-view crack-instance correspondence, but strict spatial overlap remains challenging for early-stage, weakly visible cracks.Source: Figure 10 · p. 36, §4.2 · pp. 36–37

4.3 Crack-growth tracking analysis

  • The final pipeline achieves pooled instance-level CLCE of 0.753 over 115 valid reference crack-instance transitions, with scene-level values ranging from 0.464 to 0.950.The reported sources of high CLCE include mask fragmentation, local projection drift, and few valid transitions; the pipeline captures macroscopic elongation but fine-grained length consistency remains sensitive to segmentation and registration.Source: Table 3 · p. 36, §4.3 · p. 38
  • Immediate Growth Precision is 0.759, Recall is 0.624, and F1 is 0.685; DGR@K rises from 0.303 at K=0 to 0.317 at K=2.The two-interval allowance adds 1.4 percentage points of recovered reference growth under the adopted assignment rule and available observations.Source: §4.3 · p. 39, Figure 11 · p. 39

4.4 Ablation study

  • The principal implementation parameters were held fixed across all eight scenes.The ablation metrics use a shared manually reviewed growth reference, with near-zero-growth transitions excluded from CLCE aggregation.Source: §4.4 · p. 40

4.4.1 Module-Level Ablation

  • Replacing MASt3R geometry with an ASpanFormer homography leaves HOTA(0) nearly unchanged at 0.547 versus 0.548, while CLCE rises from 0.753 to 0.769.These near-equivalent HOTA(0) values indicate that the tracking logic is not rigidly tied to one geometry backend; MASt3R yields slightly lower growth-length error under the common reference.Source: §4.4.1 · p. 40

4.4.2 Geometry Matcher Comparison

Across all evaluated transitions, MASt3R provided better downstream growth-tracking results than ASpanFormer at essentially matched valid mapping coverage. The support-band analysis found stable CLCE within 14–16 px, while error and evaluation limits remain tied to segmentation, projection, and laboratory-scale conditions.

4.4.2 Geometry Matcher Comparison

  • At essentially matched valid mapping coverage (0.774 for ASpanFormer; 0.765 for MASt3R), MASt3R achieved higher Immediate Growth F1 (0.685 versus 0.663) and lower CLCE (0.753 versus 0.769).The comparison covers all 45 evaluated transitions; valid mapping coverage is a comparability check, not a measure of projection accuracy.Source: §4.4.2 · pp. 41–42, Table 5 · p. 42

4.4.3 Sensitivity of Growth Support Band

  • CLCE increased when both growth-support radii were too small, but remained stable within 14–16 px; the default setting is 16 px for each radius.The support band absorbs projection drift and segmentation discontinuities; it does not represent physical crack width.Source: §4.4.3 · p. 42, Figure 13 · p. 43

4.5 Error propagation and practical limitations

  • Incomplete early-crack masks, residual projection error, and ambiguity in crack instantiation and association interact to impair identity tracking and growth measurement.Morphology-aware instantiation can reconnect compatible fragments but cannot recover crack support absent from the predicted mask.Source: §4.5 · p. 44
  • Evaluation covers eight laboratory scenes, one fixed segmentation checkpoint, fixed pixel-scale tolerances, and multi-instance discrimination primarily in Scenes 6–8.Field generalizability, scale transferability, and irregular inspection intervals remain unverified; future work prioritizes field validation, fine-crack segmentation, and calibrated, uncertainty-aware measurement.Source: §4.5 · pp. 44–45

5. Conclusion

The study presents a geometry-assisted, morphology-aware framework for identity-aware crack-growth monitoring under changing viewpoints, alongside an annotated dataset and evaluation measures. Experiments demonstrate quantitative crack-length characterization, while fine-crack segmentation, registration drift, and laboratory-scale evaluation limit current conclusions.

5. Conclusion

  • The framework integrates geometry-assisted alignment, morphology-aware crack instantiation, identity propagation, and pairwise growth attribution to distinguish existing crack regions from new growth without a fixed camera.It converts frame-wise crack masks into identity-consistent crack histories.Source: §5 · p. 45
  • The dataset covers Mode I opening, Mode II-dominated diagonal shear, and mixed-mode flexural–shear cracks with turning and branching; CLCE quantifies post-association crack-length growth errors.The evaluation also introduces diagnostics for same-interval growth localization and detection within a specified observation window.Source: §5 · p. 45
  • Across eight evaluated scenes, pooled instance-level CLCE was 0.753, and correspondence remained achievable under a relaxed overlap criterion.The authors report that conventional HOTA values are relatively low for crack tracking, where growth, branching, merging, and elongated morphology differ from conventional object instances.Source: §5 · pp. 45–46
  • Current limitations include fine-crack segmentation errors, residual registration drift, fixed pixel-scale tolerances, and laboratory-scale validation.The authors identify field validation, improved fine-crack extraction, adaptive physical-scale criteria, and uncertainty-aware growth quantification as future work.Source: §5 · p. 46

CRediT authorship contribution statement

The contribution statement assigns writing, methodology, project administration, funding, data curation, validation, analysis, conceptualization, and supervision across eight authors.

CRediT authorship contribution statement

  • Pengru Deng is credited with review and editing, project administration, methodology, funding acquisition, and conceptualization.Su Wang is credited with review and editing, project administration, and methodology.Source: CRediT authorship contribution statement · p. 46
  • Jiapeng Yao is credited with the original draft, data curation, validation, and formal analysis; Xinrun Li is credited with the original draft, review and editing, methodology, and formal analysis.Source: CRediT authorship contribution statement · p. 46
  • Xuhui He is credited with review and editing, funding acquisition, and supervision; Zhao Wang with review and editing and project administration.Takanori Emaru and Takashi Matsumoto are each credited with review and editing and supervision.Source: CRediT authorship contribution statement · p. 46

Related papers

Loading 2610.33550v1…