Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ENSTRECT: A Stage-based Approach to 2.5D Structural Damage Detection

Christian Benz, Volker Rodehorst

TL;DR

ENSTRECT tackles the challenge of translating 2D image-level damage detections into actionable 2.5D representations by projecting segmentations onto a 3D point cloud and extracting measurable damage instances. The framework integrates three state-of-the-art image-level detectors (TopoCrack, nnU-Net, DetectionHMA) within a three-stage pipeline—Detection, Mapping, and Extraction—to produce medial axes for cracks and PCA/alpha-shape-based polygons for areal damages. Quantitative results show IoU exceeding 90% for cracks and over 80% for corrosion at a 4 cm tolerance, while instance-level AP50 remains modest (roughly 45–56%), underscoring both the promise and current limitations of 2.5D damage quantification. The work also discusses scalability challenges of dense point clouds and points to future directions toward native 3D damage detection and improved 2D–3D fusion for practical SHM workflows.

Abstract

To effectively assess structural damage, it is essential to localize the instances of damage in the physical world of a civil structure. ENSTRECT is a stage-based approach designed to accomplish 2.5D structural damage detection. The method requires an image collection, the relative orientation, and a point cloud. Using these inputs, surface damages are segmented at the image level and then mapped into the point cloud space, resulting in a segmented point cloud. To enable further quantitative analyses, the segmented point cloud is transformed into measurable damage instances: cracks are extracted by contracting the clustered point cloud into a corresponding medial axis. For areal damages, such as spalling and corrosion, a procedure is proposed to compute the bounding polygon based on PCA and alpha shapes. With a localization tolerance of 4cm, ENSTRECT can achieve IoUs of over 90% for cracks, 82% for corrosion, and 41% for spalling. Detection at the instance level yields an AP50 of about 45% (cracks, spalling) and 56% (corrosion).

ENSTRECT: A Stage-based Approach to 2.5D Structural Damage Detection

TL;DR

ENSTRECT tackles the challenge of translating 2D image-level damage detections into actionable 2.5D representations by projecting segmentations onto a 3D point cloud and extracting measurable damage instances. The framework integrates three state-of-the-art image-level detectors (TopoCrack, nnU-Net, DetectionHMA) within a three-stage pipeline—Detection, Mapping, and Extraction—to produce medial axes for cracks and PCA/alpha-shape-based polygons for areal damages. Quantitative results show IoU exceeding 90% for cracks and over 80% for corrosion at a 4 cm tolerance, while instance-level AP50 remains modest (roughly 45–56%), underscoring both the promise and current limitations of 2.5D damage quantification. The work also discusses scalability challenges of dense point clouds and points to future directions toward native 3D damage detection and improved 2D–3D fusion for practical SHM workflows.

Abstract

To effectively assess structural damage, it is essential to localize the instances of damage in the physical world of a civil structure. ENSTRECT is a stage-based approach designed to accomplish 2.5D structural damage detection. The method requires an image collection, the relative orientation, and a point cloud. Using these inputs, surface damages are segmented at the image level and then mapped into the point cloud space, resulting in a segmented point cloud. To enable further quantitative analyses, the segmented point cloud is transformed into measurable damage instances: cracks are extracted by contracting the clustered point cloud into a corresponding medial axis. For areal damages, such as spalling and corrosion, a procedure is proposed to compute the bounding polygon based on PCA and alpha shapes. With a localization tolerance of 4cm, ENSTRECT can achieve IoUs of over 90% for cracks, 82% for corrosion, and 41% for spalling. Detection at the instance level yields an AP50 of about 45% (cracks, spalling) and 56% (corrosion).
Paper Structure (20 sections, 2 equations, 4 figures, 2 tables)

This paper contains 20 sections, 2 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Crack Segmentation
  4. Detection of Structural Damages
  5. 3D Damage Detection
  6. Dataset
  7. Enstrect
  8. Detection
  9. TopoCrack
  10. nnU-Net
  11. DetectionHMA
  12. Mapping
  13. Extraction
  14. Cracks
  15. Areal damages
  16. ...and 5 more sections

Figures (4)

  • Figure 1: Components and workflow of the 2.5D detection pipeline.
  • Figure 2: Illustration of the cloud contraction for extracting the medial axis of a branching crack: (a) shows a mesh with high-resolution texture, (b) the subcloud after clustering, (c) the overlayed medial axis (red) obtained by cloud contraction, and (d) the medial axis overlayed over the textured mesh.
  • Figure 3: Illustration of the extraction of the bounding polygon for areal damages: (a) shows a textured mesh with spalling and corrosion, (b) the segmented point cloud, (c) the corresponding bounding polygon in 2.5D space around the clustered point cloud (spalling and corrosion are merged in one cluster), and (d) the bounding polygon overlayed on the textured mesh.
  • Figure 4: Qualitative test results for 2.5D damage detection. The top row shows the test segment of Bridge B, the bottom row shows the test segment of Bridge G. The results of nnU-Net isensee2021nnu and DetectionHMA benz2022image are compared. While nnU-Net achieves better results for spalling and corrosion, DetectionHMA benz2022image shows more robust performance for crack detection. In zoomed view: medial axis and bounding polygons overlayed on textured mesh. Best viewed on screen.