Open AccessDeep Fusion and Ensemble Neural Networks for Point Sets Data
Author(s) -
Yi Kang,
Zhang Na,
Yuting Zhao,
Yiyan Lei
Publication year - 2019
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/1325/1/012097
Subject(s) - computer science , point cloud , artificial neural network , artificial intelligence , segmentation , point (geometry) , pattern recognition (psychology) , representation (politics) , feature (linguistics) , object (grammar) , permutation (music) , noise (video) , data mining , image (mathematics) , mathematics , linguistics , philosophy , physics , geometry , politics , political science , acoustics , law
The typical character of point cloud is that its format is irregular, so most researchers transform such data into regular format. However, this conversion may induce some noise. In this paper, we design the dedicated neural network unit for point sets data processing. Firstly, to guarantee the permutation invariance of point sets, we infer the neural unit design principles. Secondly, we propose the deep fusion and ensemble neural networks for point sets data, in which we can not only combine different scales of feature representation, but also input the point data into the network directly. Finally, we evaluate the proposed networks on object classification and object part segmentation, which can verify the efficiency of our network.