Semantic Segmentation Network with Adaptive Neighborhood Construction for 3D Point Cloud

Abstract

In the process of point cloud feature extraction, the inconsistent local spatial density of the input points will lead to the instability of Information fusion in the neighborhoods. In this paper, an adaptive neighborhood construction algorithm is proposed to solve the problem of uneven density when constructing spatial neighborhood from point cloud, which can more accurately capture the local spatial geometry near the query point. In addition, we augment the attentive pooling with differential attentive function, and both are integrated to assign a more effective attention score to the encoded feature. We applied the above methods to the RandLA-Net and tested on S3DIS Area5. The experiments show that our methods can effectively optimize the result of semantic segmentation and have obvious advantages in terms of accuracy and robustness.