Semantic understanding of point clouds based on density-adaptive sampling and position-sensitive feature extraction

Semantic understanding of point clouds is a major task in 3D scenes. Due to the enormous number of points within point clouds, it is strait to process point clouds directly. Generally, previous work follow a sampler-extractor structure to get high-level semantic features. Previous sampling strategies in the sampler stuck in upholding local density while guaranteeing a predetermined number of sampled points. We propose a density-adaptive sampling strategy. In the extractor, PointNN introduces a non-parameter framework. However, it ignores the positional distance among points. We introduce a position-sensitive feature extraction module inspired by PointNN. We conduct comprehensive experiments to validate the effectiveness of our method.