Dual-Neighborhood Deep Fusion Network for Point Cloud Analysis

• DOI: 10.48550/arxiv.2108.09228
• Publication Date: 2021-01-01
• Authors: Guoquan Xu; Hezhi Cao; Yifan Zhang; Jianwei Wan; Ke Xu; Yanxin Ma
• Keywords: Transfer of learning; Representation; Feature Learning; Feature (linguistics); Convolution (computer science); Fusion mechanism

Recently, deep neural networks have made remarkable achievements in 3D point cloud classification. However, existing classification methods are mainly implemented on idealized clouds and suffer heavy degradation of per-formance non-idealized scenarios. To handle this prob-lem, a feature representation learning method, named Dual-Neighborhood Deep Fusion Network (DNDFN), is proposed to serve as an improved encoder for the task DNDFN utilizes trainable neighborhood method called TN-Learning capture global key neighborhood. Then, fused with local neighbor-hood help network achieve more powerful reasoning ability. Besides, Information Transfer Convolution (IT-Conv) learn edge infor-mation between point-pairs benefits transfer procedure. The transmission information IT-Conv similar propagation graph which makes closer human mode. Extensive experiments benchmarks especially datasets verify effectiveness achieves state arts.