MASSTAR is a multimodal large-scale scene dataset jointly proposed by Sun Yat-sen University, Hong Kong University of Science and Technology and other institutions, focusing on surface prediction and completion tasks. The dataset contains more than 1,000 scene-level 3D mesh models, some of which are from the real world. In addition to 3D models, the dataset also contains corresponding images, descriptive text, and point cloud data.

To create the MASSTAR dataset, researchers developed a versatile tool chain that can efficiently extract high-quality 3D mesh models from complex scenes and generate corresponding multimodal data. MASSTAR has broad application prospects in robotics, high-quality 3D reconstruction, and autonomous driving, and provides strong data support for solving surface completion problems in complex scenes.