VE444 Networks

Group13 Course Project

Deformable Graph Matching for 2D Point Cloud

Network Analysis
Team Members: Kunyi Yang, Jiayue Ding
Instructor: Prof. Yifei Zhu

Overview Algorithm Design Results

• Graph matching is crucial in computer The factorization of the 𝑲𝑲 affinity matrix Dataset: CMU House Dataset (vertices of a
science. proposed in the DGM paper is: rotating toy house, 111 images). The
• Traditional methods struggle with 𝑲𝑲 = 𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑 𝑣𝑣𝑣𝑣𝑣𝑣 𝑲𝑲𝑝𝑝 vertices are connected (undirectedly) using
scalability and robustness. +(𝑮𝑮2 ⊗ 𝑮𝑮1 ) 𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑 𝑣𝑣𝑣𝑣𝑣𝑣 𝑲𝑲𝑞𝑞 (𝑯𝑯2 ⊗ 𝑯𝑯1 )′
Delaunay triangulation.
• This project implements Deformable
Which can then be used to simplify the
Graph Matching (DGM) in MATLAB.
optimization problem Jdgm :
• Evaluated on the CMU House dataset.
max 𝐽𝐽𝑔𝑔m 𝑿𝑿
Motivation X,𝒯𝒯

• Traditional graph matching struggles = 𝑡𝑡𝑡𝑡 𝑲𝑲𝑇𝑇𝑝𝑝 𝑿𝑿 + λ𝑡𝑡𝑡𝑡(𝑲𝑲𝑇𝑇𝑞𝑞 (𝐺𝐺1𝑇𝑇 𝑿𝑿𝐺𝐺2 ∘ 𝐻𝐻1𝑇𝑇 𝑿𝑿𝐻𝐻2 ))
with scalability and robustness, Previous works on GM explicitly computed
especially in dynamic networks. the computationally expensive K of size
• Realistic networks exhibit structural 𝑂𝑂(𝑛𝑛12 𝑛𝑛22 ). In DGM the K is calculated using
variations due to evolving connections the above equation.
and user behavior. The ultimate goal of GM algorithm is to
maximize Jdgm . In this work, 2 optimization
Problem Statement methods (CCCP and FW) are combined in
Goal: Find the optimal correspondence
one structure named “path following” to
between nodes of two graphs, maximizing
guarantee efficient convergence.
structural and attribute similarity.
max Jdgm 𝑿𝑿, 𝒯𝒯
X,𝒯𝒯
= vec 𝑿𝑿 T 𝑲𝑲vec(𝑿𝑿)
= tr 𝑲𝑲p 𝒯𝒯 T 𝑿𝑿
+ λtr 𝑲𝑲q 𝒯𝒯 T 𝐺𝐺1𝑇𝑇 𝑿𝑿𝐺𝐺2 ∘ 𝐻𝐻1𝑇𝑇 𝑿𝑿𝐻𝐻2 − ψ 𝒯𝒯
Correspondence
X
matrix Evaluation Metrics:
• Matching Accuracy: The percentage of
𝒯𝒯 Transformation
correctly matched nodes.
Jdgm Objective function • Speed: The time to compute matrix 𝑲𝑲
using the traditional / new method [1].
Transformation
𝒯𝒯
(optional) Evaluation results:
• The DGM algorithm outperform 2
Kp Node affinity matrix
other algorithms in comparison, and is
Kq Edge affinity matrix Fig. 2: pseudocode for path following DGM less prone to irrelevant outlier nodes.
Parameter to Compared Method • The factorization approach to compute
the matrix 𝑲𝑲 is significantly faster than
λ balance node and • Graduated Assignment: GA performs
edge consistency. the traditional method.
gradient ascent on a relaxed objective
Constraint function Findings:
function driven by annealing schedule.
that penalizes • The graph matching method naturally
ψ 𝒯𝒯 At each step, it maximizes a Taylor
complex cannot distinguish image depth because
transformations
expansion of the non-convex object
it only considers the structure of the
function around the previous
tr Trace of a matrix graph and the position of the nodes.
approximate solution.
• Spectral Matching: SM optimizes a References
relaxed quadratic programming that 1. Zhou, Feng, and Fernando De la Torre.
drops the affine constraints and "Deformable graph matching." Proceedings
of the IEEE Conference on Computer Vision
introduces a unit-length constraint on 𝑿𝑿.
and Pattern Recognition, 2013.
2. S. Gold and A. Rangarajan, "A graduated
assignment algorithm for graph matching,"
in IEEE Transactions on Pattern Analysis and
Machine Intelligence, vol. 18, no. 4, pp. 377-
388, April 1996.
3. A. Egozi, Y. Keller and H. Guterman, "A
Fig. 1: Graph Matching Problem Probabilistic Approach to Spectral Graph
Matching," in IEEE Transactions on Pattern
Analysis and Machine Intelligence, vol. 35,
no. 1, pp. 18-27, Jan. 2013.

University of Michigan - Shanghai Jiao Tong University Joint Institute