Concrete400: A multi-task pixel-level crack dataset from urban infrastructure in Vietnam

Concrete400 is a multi-task concrete crack dataset for visual infrastructure inspection. It contains 379 RGB photographs of building surfaces and concrete pavements in Hanoi, Vietnam, with manually prepared pixel-level crack masks and corresponding text files containing crack region bounding box annotations. The release also includes 1502 labelled image patches for binary crack classification, comprising 633 crack patches and 869 no-crack patches. The collection preserves field variability in illumination, concrete texture, surface condition, background clutter, and acquisition time. The released segmentation and object detection images, masks, and box coordinates are standardised to 512 $\times$ 512 pixels, while the classification patches are provided at 256 $\times$ 256 pixels. The accompanying reproducible protocol defines 303 training and 76 validation image-mask pairs. After masks are binarised at an intensity threshold of 127, crack pixels represent 3.08% of image pixels on average. Concrete400 can therefore support patch-level crack classification, semantic crack segmentation, instance-style crack region segmentation, object detection, semi-supervised and self-supervised learning, and crack image generation under realistic inspection conditions.