ABSTRACT: In this post we are discussing the
similarities of graphs and images. We will start with matrix and spacial
representations to show how graphs and images differs and resembles.
Then we will dive into the applications from traditional signal
processing to neural networks. Finally we take 3D vision as the bridging
of image and graph in computer vision tasks.
Draft structure:
Basics (Spacial representation)
Matrix representation of graph
Non-Euclid properties of graph
Matrix representation of image
Euclid properties of images
Image is special graph
From group theory's perspective
Applications
Signal and systems review, but 2d
Frequency domain of graph and img
Sampling, filtering and convolution of graphs and
images.
CNN & GNN (Local receptive fields in CNNs vs. neighborhood
aggregation in GNNs)
Graph transformers & ViT
Combinational application
PointNet Series
Draft end.
Spacial Representations
Graphs
A graph consists of
vertices and
edges . Here we have a
denotes the signal assigned to
each node, it can be a scaler, vector, or even matrix. When we deal with
graph signals, besides the signal on each node, the topology of graph
should also be considered. To record such topology, we bring up the
LAPLACIAN MATRIX.
DEFINITION. Each element of laplacian matrix is
defined as:
DEFINITION. Symmetric normalized laplacian is given as:
𝕚𝕗𝕚𝕗𝕠𝕥𝕙𝕖𝕣𝕨𝕚𝕤𝕖
The non-Euclid properties
of Graphs
THEOREM. A graph if and only if it can be embedded on the surface of
a sphere.