PGN3DCD: Prior-Knowledge-Guided Network for Urban 3-D Point Cloud Change Detection

Efficient and intelligent 3-D change detection is a fundamental problem in urban data updating, management, and planning. However, this paradigm faces some challenges: 1) there is a lack of publicly available datasets of realistic urban 3-D change detection and 2) the long-tailed distribution of 3-D variation is intractable for existing 3-D change detection methods. To address these issues, we present the first urban-scale realistic point cloud change detection dataset from Hong Kong with nearly 128 million binary-change annotated points, called the Hong Kong change detection (HKCD) dataset, which is the largest realistic point cloud change detection dataset. This dataset consists of photogrammetric point clouds from Hong Kong, covering about 8.1 km2 of the city landscape. Moreover, we propose a novel prior-knowledge-guided network for 3-D point cloud change detection (PGN3DCD). Specifically, we build a nonparametric and generalizable 3-D change detection prior knowledge, which guides the neural network to focus on the changes, thus more accurately characterizing the variations. Experimental results on several 3-D change detection benchmarks show a significant improvement, demonstrating the innovation and sophistication of PGN3DCD for 3-D change detection.