[2P5-GS-3] Data mining: Fundamental theory

[2P5-GS-3-03] A PointNet-based CNN Framework for Spatial Data Interpolation

〇Koh Takeuchi¹, Hisashi Kashima^1,2, Naonori Ueda^2,3 (1. Kyoto University, 2. RIKEN AIP, 3. NTT)

Keywords:Spatio-Temporal Data Analysis, PointNet

The recent advance of Convolutional Neural Network has achieved a significant improvement in the performance of missing value interpolation problems, including the super-resolution of low-resolution images and the upsampling of low-frequency audio. Despite such succeeds, CNN base methods have not been applied to the spatial interpolation problem because observing locations of spatial data do not have a regular shape. In this paper, we propose a new CNN framework based on the rising techniques based on PointNet that has been proposed to tackle pattern recognition tasks with point cloud data. We conduct preliminary experiments with a bike-sharing data set and demonstrate that our method showed a significant improvement from existing baselines.