Omni-Crack30k: A Benchmark for Crack Segmentation and the Reasonable Effectiveness of Transfer Learning

Abstract

Large benchmarking datasets, such as ImageNet, COCO, Cityscapes, or ScanNet, have enormously promoted research in computer vision. For the domain of crack segmentation, no such large and well-maintained benchmark exists. Crack segmentation is characterized by the decentralized creation of datasets, almost all of which have their specific right to existence. Each dataset covers a different aspect of the surprisingly complex landscape of materials, acquisition conditions, and appearances linked to crack segmentation. The OmniCrack30k dataset forms the first large-scale, systematic, and thorough approach to provide a sustainable basis for tracking methodical progress in the field of crack segmentation. It contains 30k samples from over 20 datasets summing up to 9 billion pixels in total. Featuring materials as diverse as asphalt, ceramic, concrete, masonry, and steel, it paves the road towards universal crack segmentation, a currently under-explored topic. Experiments indicate the effectiveness of transfer learning for crack segmentation: nnU-Net achieves a mean clIoU4px of 64% outperforming all other approaches by at least 10% points.