Meng
Tang,
PhD
candidate
David
R.
Cheriton
School
of
Computer
Science
Convolutional neural networks have been a great success for computer vision tasks including classification, segmentation, and pose estimation, etc. However, CNNs cannot be directly applied to irregular data such as point clouds or graphs, which can be obtained from the nowadays ubiquitous depth sensor.
This talk reviews recent deep learning techniques for such irregular data of sets and graphs, and shows several applications in computer vision and computer graphics. In particular, we will discuss graph neural network (GNN), which can be constructed in the spatial domain or the spectral domain. The two methods of spatial analysis and spectral analysis are highly related, and even equivalent in some cases. We will introduce graph convolution based on graph Fourier transform in the spectral domain. We will also discuss how graph neural network is related to mean-field inference for MRF. In this seminar, we focus on 3D segmentation as an example to show how GNN can be utilized.