The Campus3D is a photogrammetry point cloud dataset collected by UAV imagery over the National University of Singapore (NUS). The total dataset covers the campus of 1.58 km2 area with about 937 million points. To facilitate the research of 3D scene understanding, we also annotated such data with instance-based and hierarchical semantic labels.
In more details, point cloud of our dataset is divided into six regions: FOE, FASS, Utown, PGP, UCC and Ridge area based on their locations, functions and architectural features. We also present statistics and distribution of each region as an interactive map in Description.
Our dataset annotates the point cloud into 23 hierarchical categories and 2,530 instances, which are presented in the format of tree graph in Description. For the convenience of dataset usage, we provide several tools at Github (shinke-li/Campus3D):
a) Format transformation and label modification, which allows the referrers to flexibly merge categories for different tasks.
b) Data sampler for scene understanding, including KNN sampler and Block sampler.
c) Reduced version of Campus3D, since the memory and computation cost are heavy for learning on origin Campus3D.
Overall, we are aiming at benchmarking learning methods to various 3D scene understanding tasks containing semantic and instance segmentation. This will make contributions for research progress in various areas like computer vision, geographic information system, computer graphics and drive new solutions and problems in related areas such as 3D model reconstruction, urban planning, autonomous driving and etc.