RGB + Depth + Point Cloud + Labels Cracks

This study constructs a multimodal dataset for crack perception, where an Intel RealSense L515 LiDAR sensor is used to synchronously acquire RGB images, depth information, and three-dimensional point cloud data in three representative scenarios, namely concrete cracks on building fa\u00e7ades, surrounding-rock cracks in underground coal mine roadways, and pavement cracks. After extrinsic and intrinsic calibration, the point clouds are accurately projected onto two-dimensional planes with the same resolution as the RGB images to generate geometrically aligned auxiliary channels, thereby enabling the joint representation of texture and geometric cues. Each sample in the dataset contains registered RGB images, depth maps, and point cloud data together with pixel-level crack annotations and JSON metadata, covering diverse material types, illumination conditions, surface contamination, and occlusion patterns. In total, the dataset comprises 2,858 samples across the three scenarios and provides a unified and realistic benchmark for the training, validation, and cross-scene generalization evaluation of multimodal crack segmentation networks.