Beyond Cracks: Synthetic Image and Geometry Generation for Computer Vision Detection and Severity Assessment of Diverse Concrete Surface Defects

Shun-Hsiang Hsua, Mani Golparvar-Farda,b
aDepartment of Civil and Environmental Engineering, bSchool of Computing and Data Science
University of Illinois at Urbana-Champaign
Under Review, Automation in Construction, 2026
Paper Code Datasets
Pipeline overview: 2D defect shapes to 3D parametric models to multi-pass rendered outputs

Overview of the geometry-informed synthetic data generation pipeline. Real 2D defect shapes are transformed into parametric 3D models and rendered with domain randomization to produce aligned RGB, depth, normal, and mask outputs.

Abstract

Computer vision-based concrete condition assessment has largely focused on crack detection, as existing datasets lack multi-defect coverage and geometry-grounded severity labels. A two-fold contribution is presented. First, a geometry-informed synthetic image generation pipeline is developed to convert real defect shapes into parametric 3D models, supporting scalable rendering of diverse defects with paired RGB, depth, and surface-normal modalities. The pipeline is extensible and supports future modeling needs beyond the defect classes considered in this work. Second, a large-scale synthetic dataset with automatic annotations is established, jointly encoding semantic defect categories and geometry-informed cues, facilitating severity-aware training and evaluation across cracks, spalling, corrosion, and efflorescence. Building on this data, this paper proposes a geometry-informed teacher--student framework that leverages geometric modalities during training while maintaining RGB-only inference, allowing the application to both close-range crack inspection and structural-scale assessment. To our knowledge, this represents the first scalable framework to unify real-shape-grounded synthetic generation, geometry-informed learning, and severity-oriented concrete condition assessment, while enabling systematic training and benchmarking for non-crack defects.

Method

Cube-based Crack Modeling

For close-range crack inspection, a six-surface cube serves as the carrier. Crack masks from a curated real-image database are placed on cube faces with parameterized scale, orientation, and position. Pixel-to-metric conversion combined with FHWA thresholds assigns crack severity (insignificant, moderate, wide), and crack geometry is modeled with configurable width, erosion depth, and overall depth for 3D realism.

Cube-based crack modeling process
Cube-based crack modeling: crack masks placed on cube surfaces, converted to 3D geometry with severity assignment.

Component-based Bridge Defect Modeling

For structural-scale bridge inspection, parameterized bridge elements (deck, pier, girder, parapet, bearing) are generated from road centerlines and structural guidelines. Defect masks for crack, spalling, corrosion, and efflorescence are instantiated on component surfaces. Category-specific 3D operations model surface erosion (spalling), exposed reinforcement (corrosion), and extruded deposits (efflorescence), integrating defect and structural geometry in a unified process.

Component-based bridge defect modeling
Component-based bridge modeling with defect placement and domain randomization for rendering.

Datasets

All synthetic datasets and a curated real-image evaluation set are prepared for public release.

Dataset Images RGB Depth Normal Mask Description
SynthCrack-42k 42,165 ✓ ✓ Baseline cube-based crack dataset (Unreal Engine) Download
SynthCrack-72k 71,860 ✓ ✓ Extended crack shapes + updated geometric parameterization Download
SynthCrack-ONE 23,485 ✓ ✓ ✓ ✓ Rhino-generated with refined crack-shape library Download
SynthDefect-Bridge 2,291 ✓ ✓ ✓ ✓ Bridge-scale multi-defect: crack, spalling, corrosion, efflorescence Download
RealDefect-Bridge 2,420 ✓ ✓ Real bridge inspection images for evaluation Download

Results

Multi-pass Outputs

Each scene produces aligned color, depth, surface normal, and defect mask outputs from the same camera pose.

Color output

Color

Depth output

Depth

Normal output

Normal

Mask output

Mask

Realism Enhancement

Unpaired image-to-image translation (CycleGAN, CycleGAN-Turbo, CUT) reduces the appearance gap between synthetic and real inspection imagery.

Qualitative comparison of realism enhancement methods for close-range crack imagery
Realism enhancement comparison for close-range crack imagery.
Qualitative comparison of realism enhancement methods for structural-scale defect scenes
Realism enhancement comparison for structural-scale bridge defect scenes.

Defect Predictions on Real Images

Models trained on synthetic data transfer to real inspection imagery for crack severity prediction and multi-defect semantic segmentation.

Crack severity predictions on real images
Crack segmentation with class-conditioned severity prediction on real images.
Bridge defect semantic segmentation predictions on real images
Multi-defect semantic segmentation on real bridge inspection images.

BibTeX

@article{hsu2026defectsynth,
  title={Beyond Cracks: Synthetic Image and Geometry Generation for Computer Vision Detection and Severity Assessment of Diverse Concrete Surface Defects},
  author={Hsu, Shun-Hsiang and Golparvar-Fard, Mani},
  journal={Automation in Construction},
  year={2026},
  note={Under Review},
  url={https://huhuman.github.io/rhino-defect-synth/}
}

Acknowledgements

This material is based upon work supported by the National Science Foundation (NSF) under Grant No. CMMI-2053935. This work used the Delta advanced computing resource, supported by the NSF (OAC 2005572) and the State of Illinois, a joint effort of the University of Illinois and the National Center for Supercomputing Applications (NCSA). Any findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the NSF or NCSA.