"CrackBlock50k"

"CrackBlock50K is a large-scale, high-resolution pavement crack dataset constructed from 50,000 real-world road images collected in highway scenarios. The dataset is designed to support robust crack segmentation under noisy supervision, with annotations provided in a structured block-level format to balance labeling efficiency and spatial precision. To further enhance training effectiveness and local crack representation, an overlapping cropping strategy is applied to generate over 600,000 patch-level samples from the original images. These derived samples improve data diversity, strengthen fine-grained feature learning, and alleviate class imbalance without introducing additional annotation cost. By combining large-scale real data with an efficient patch-based representation, CrackBlock50K provides a practical and scalable benchmark for both crack segmentation and noise-robust learning in intelligent transportation systems."