CrackInstSynth: Topology-Aware Generative Data-Augmentation Framework for Crack Instance Segmentation

ICLR 2026 Conference Submission20577 Authors

Keywords: crack instance segmentation, topology-aware diffusion, generative data augmentation, structural health monitoring

Abstract:

Instance-level crack segmentation is critical for automated structural health monitoring of tunnels and bridges, yet progress is limited by the scarcity of densely annotated datasets with instance-level labels.
To address this gap, we make two key contributions. First, we introduce CrackInst1K, to our knowledge the first publicly available instance-level crack segmentation dataset, comprising 1025 high-resolution tunnel images with pixel-accurate instance masks. Second, we propose CrackInstSynth, a generative data-augmentation framework that substantially enlarges instance-level crack corpora while preserving geometric and topological realism. CrackInstSynth integrates three coordinated modules: (i) Region-level Instance Placement (RIP), which partitions the canvas into quadrants to strategically position crack instances for diverse layouts; (ii) a Physics-driven Skeleton Generator (PSG) that enriches morphological variability by growing crack skeletons via physical simulation; and (iii) a Topology-Preserving Generation Module (TPGM) that employs a two-stage conditional diffusion pipeline (skeleton→mask, mask→image) to produce paired, width-aware instance masks and corresponding images while enforcing intra-instance topology and inter-instance separation. Extensive experiments show that augmenting real data with CrackInstSynth consistently improves the performance of multiple instance segmentation models on CrackInst1K and other benchmarks, validating both visual fidelity and downstream effectiveness. We will release CrackInst1K and CrackInstSynth to facilitate future research in structural health monitoring.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 20577