2Dinsight3D: an multi-modal urban scene dataset for point cloud semantic segmentation from remote sensing imagery

Inspired by image domain adaptation and projection-based point cloud segmentation methods, we propose a new dataset, 2Dinsight3D, and an efficient method, labels adapt 3D (LA3D), for the cost-effective point cloud semantic segmentation. First, the 2Dinsight3D dataset provides an image-point cloud alignment dataset, designed to evaluate point cloud segmentation using image labels as ground truth. Next, the dataset offers several key advantages: (1) it includes diverse environments such as urban areas, mountains, and ports; (2) the point cloud data are enriched with RGB feature information alongside 3D coordinates; and (3) both point cloud and corresponding remote‑sensing imagery datasets are fully labeled. Based on this, the LA3D method is introduced, which generates labels from remote‑sensing imagery labels to point cloud labels. This approach significantly reduces labeling costs by relying on imagery labels. Experimental results demonstrate that the performance using coarse-grained pseudo‑imagery labels is comparable to that achieved with ground truth point cloud labels.