• • View
  • Edit
  • Attach
  • Print

Graph-based CNN for Human Action Recognition from 3D Pose

Mike Edwards and Xianghua Xie

Abstract
Deep learning has shown increasingly promising performance in the pattern recognition field in recent years, becoming a prominent staple in image classification problems since the introduction of the convolutional neural network. Such CNN models work well in the image domain due to the spatially regular structure of the 2D and 3D grid, but not all domain exhibit such a regular spatial structure. In order to retain the underlying spatial information within the domain application, this study presents operators for graph-based convolution and pooling, utilizing graph based signal processing methods to define common deep learning operators, such as convolution and pooling, on a graph representation of the spatial human skeleton domain. The proposed method avoids unnecessary assumptions of spatial relationships between hand-crafted features, and evaluation shows strong sequence classification rates that exceeds 93%.

Files
Paper (PDF)

DOI
Coming soon

Bibtex
@inproceedings{dlid2017_1,

}