synth-dacl: Synthetic Extensions for dacl10k Dataset

Adequate bridge inspection is increasingly challenging due to aging infrastructure, a growing number of structures requiring attention, and limited personnel and financial resources. Automating key inspection tasks — such as the pixel-level classification of defects and structural components — relies on robust datasets that reflect real-world diversity and complexity. The dacl10k dataset has emerged as the most comprehensive resource for visual inspection of concrete bridges, offering a wide variety of annotated images captured under different environmental conditions, camera settings, and structural contexts. However, it suffers from class imbalance, particularly in underrepresented defect classes like cracks and cavities.
To address this limitation, we introduce synth-dacl, a suite of three synthetic dataset extensions — daclonsynth (semi-synthetic), synthcrack, and synthcavity — generated using diverse synthesized concrete textures. Each extension targets specific limitations of dacl10k: (i) daclonsynth is created by pasting cropped defect regions from dacl10k onto synthetic concrete backgrounds, (ii) synthcrack features procedurally generated cracks applied to synthetic surfaces, and (iii) synthcavity contains synthetic cavities rendered on artificial concrete textures.
These extensions are explicitly designed to balance class distributions and enrich the visual variability of defect representations, thereby facilitating the training of more robust and generalizable models for future research in automated bridge inspection.