CCIC-O (Concrete Crack Images for Classification - Orientation)

CCIC-O (Concrete Crack Images for Classification - Orientation) is a derived crack dataset constructed from the base dataset introduced by Özgenel and Sorguç [1] and released by Özgenel as Concrete Crack Images for Classification (CCIC) [2]. It is designed for fine-grained ordinal evaluation by providing controlled variation in crack orientation. In addition to the rotated images, the dataset contains the 1,008 un-rotated source cracks with their measured, continuous orientation labels (natural configuration). Source images are filtered through an automated image-processing pipeline to retain cases dominated by a single, approximately straight, edge-to-edge crack. The remaining images are aligned to a common horizontal reference orientation, manually validated, and then rotated in 10° increments over the full circle, yielding 36 orientation classes and 36,288 images (1,008 source cracks × 36 rotations) at 227 × 227 pixels. To avoid dark border artifacts introduced by rotation, images are reflectively padded before rotation and center-cropped back to the target resolution afterwards. Since a crack line is symmetric under 180° rotation, the 36 physical orientations correspond to 18 effective orientation classes (modulo 180°). The folder natural/ contains the source cracks in their original, un-rotated form, unchanged from CCIC. The accompanying file labels_natural.csv maps each image to the continuous orientation measured during the alignment step. This natural configuration complements the balanced, discrete rotated classes with fully real images and continuous labels. [1] Özgenel, Ç.F., Sorguç, A.G. Performance Comparison of Pretrained Convolutional Neural Networks on Crack Detection in Buildings. ISARC 2018, Berlin, 2018. [2] Özgenel, Ç.F. Concrete Crack Images for Classification. Mendeley Data, V2, 2019. DOI: 10.17632/5y9wdsg2zt.2