ProbabilisticGraph and Hypergraph Matching
Ron Zass & Amnon Shashua
School of Engineering and Computer Science,The Hebrew University, Jerusalem
Zass & Shashua
Example: Object Matching No global affine transform.
Images from: www.operationdoubles.com/one_handed_backhand_tennis.htm
Local affine transforms + small non-rigid motion.
Match by local features + local structure.
Zass & Shashua
Hypergraph Matching inComputer Vision
Problem: Distances are not affine invariant.
In graph matching, we describe objects asgraphs (features nodes, distances edges)and match objects by matching graphs.
Zass & Shashua
Hypergraph Matching inComputer Vision
Affine invariant properties. Properties of four or more points. Example: area ratio, Area1 / Area2 Describe objects as hypergraphs
(features nodes, area ratio hyperedges)
Match objects by doing Hypergraph Matching.
In general, if n points are required to solve the local transformation, d = n+1 points are required for an invariant property.
2
13
4
Area1
Area2
Zass & Shashua
Related WorkHypergraph matching
Hypergraph matching: Wong, Lu & Rioux, PAMI 1989 Sabata & Aggarwal, CVIU 1996 Demko, GbR 1998 Bunke, Dickinson & Kraetzl, ICIAP 2005
All search for an exact matching. Edges are matched to edges of the exact same
label. Search algorithms for the largest sub-isomorphism.
We are interested in an inexact matching. Edges are matched to edges with similar labels. Find the best matching according to some score
function.
Unrealistic
Zass & Shashua
Related WorkInexact Graph Matching
Popular line of works: Continuous relaxation As an SDP problem
Schellewald & Schnörr, CVPR 2005 As a spectral decomposition problem
Leordeanu & Hebert, ICCV 2005; Cour, Srinivasan & Shi, NIPS 2006 Iterative Linear approximations using Taylor
expansions Gold & Rangarajan, CVPR 1995
And many more. Some continuous relaxation may be interpreted
as soft matching.
Our work differ:We assume probabilistic interpretation of
the input and extract probabilistic matching
Zass & Shashua
From Soft to Hard
Given the optimal soft solution X , the nearest hard matching is found by solvinga Linear Assignment Problem. The two steps (soft matching and nearest
hard matching) are optimal. The overall
hard matchingis not optimal(NP-hard).
SoftMatching
HardMatching
BetterMatching
Zass & Shashua
Hypergraph Matching
Two directed hypergraphs of degree d,G=(V,E) and G ’=(V ’,E ’). A hyper-edge is an ordered d -tuple of
vertices. Include the undirected version as a private
case. Matching: m : V V ’ Induce edge matching, m : E E ’,
m (v1 ,…,vd ) = (m (v1 ),…,m (vd ) )
Zass & Shashua
ProbabilisticHypergraph Matching
Input: Probability that an edge frome E match to an edge in e ’ E ’:
Output: Probability that two vertices match
We will derive an algebraic connection between S and X , and then use it for finding the optimal X .
',|')(Pr', GGeemS ee
',|')(Pr', GGvvmX vv
Zass & Shashua
Kronecker Product
Kronecker product between an i xj matrix A to a k xl matrix B is a ik xjl matrix:
BaBa
BaBaBA
ijj
i
1
111
didi AAA 11
AA di
d1
Zass & Shashua
S ↔ X connection
Assumption Proposition (S ↔ X connection)
Proof
',|)()( 21 GGvmvm
XS d
d
ivv
d
iii
ee
iiX
GGvvm
GGeemS
1',
1
',
',|')(Pr
',|')(Pr
',,''
,,,
1
1
d
d
vvevve
Zass & Shashua
S ↔ X connection for graphs
V ’ xV ’
V xV
XX S
V ’
V X
Xxij
XXS
',',)','(),,( 22112121 vvvvvvvv XXS
Zass & Shashua
Globally Optimal Soft Hypergraph Matching
Nearest to S , where X is a validmatrix of probabilities:
Vertex can be left unmatched. With equalities, all vertices must be
matched.
1111
T
d
X
XXXts
XSdist
,,0..
,min
Xd
1111 TXX ,
Zass & Shashua
Cour, Srinivasan & Shi 2006
Our result can explain somepreviously used heuristics.
Cour et al 2006 preprocessing:Replace S with the nearest doubly stochastic matrix (in relative entropy) before any other graph matching algorithm.
Proposition: For X ≥ 0, X is doubly stochastic iff is doubly stochastic.
is doubly stochastic.Xd
XS d
Zass & Shashua
Globally Optimal Soft Hypergraph Matching
We use the Relative Entropy (Maximum Likelihood) error measure,
Global Optimum, Efficient.
1111
T
d
X
XXXts
XSdist
,,0..
,min
ijijji ij
ijij BA
B
AABADBAdist
,
log)||(),(
Zass & Shashua
Globally Optimal Soft Hypergraph Matching
Define:
XSDX d
X ||minarg*
1111 XXXYDX TdT
X ||minarg*
d
iveevee
eevv
i
i
SY1
'''||
',',
Convex problem, with |V |x|V ’| inputs and outputs!
Zass & Shashua
Globally Optimal Soft Hypergraph Matching
Define , the number of matches.
1111 XXXYDX TdT
X ||minarg*
11 Xk T
kXXXX
tsXYDkX
TTX
111111 ,,,0
..||minarg*
X *
(k ) is convex in k. We give optimal solution for X
* (k ),
and solve for k numerically(convex minimization in single variable).
Zass & Shashua
Globally Optimal Soft Hypergraph Matching
Define three sub-problems ( j =1,2,3):
Each has an optimal closed form solution.
321..||min CCCXtsXYDX
kXXXC
XXXC
XXXC
T
T
11
11
11
,0|
,0|
,0|
3
2
1
)(||min)( HXtraceXYDHP T
CXj
j
Zass & Shashua
Successive Projections[Tseng 93, Censor & Reich 98]
Set For t =1,2,… till convergence:
For j = 1,2,3:
.,0 )0()0( YX jj
)(1
)1()( tj
tjj
tj XfPX
)()(1
)1()( tj
tj
tj
tj XfXf
XYDXfXX tt ||,)1(3
)(0
Optimal!
Zass & Shashua
Globally Optimal Soft Hypergraph Matching
When the hypergraphs are of the same size, and all vertices has to be matched,
our algorithm reduces to the Sinkhorn algorithm for nearest doubly stochastic matrix in relative entropy.
,111 TXX
Zass & Shashua
Sampling Given Y, the problem size reduce to |V |x|V ’|. Calculate Y : simple sum on all hyper-edges. Problem: Compute S, the hyper-edge
to hyper-edge correlation. Sampling heuristic: For each vertex, use
only z closest hyper-edges.
Heuristic applies to transformation that are locally affine (but globally not affine).
O(|V |·|V ’|·z 2) correlations.
Zass & Shashua
Runtime
Spectral MatchingLeordeanu05
Our scheme (graphs)
Our scheme (hypergraphs)
Without edge correlations time With hyperedge correlations time
(50 points)
Hyperedges per vertex
Zass & Shashua
Experiments on Graphs
to a single graph to both graphs
Spectral MatchingLeordeanu05 withCour06
preprocessing
Spectral MatchingLeordeanu05
Our scheme
Two duplicates of 25 points.
Graphs based on distances.
Additional random points.
Zass & Shashua
Experiments on Graphs
Spectral MatchingLeordeanu05 withCour06
preprocessing
Spectral MatchingLeordeanu05
Our scheme
Mean distance between neighboring points is 1.
One duplicate distorted with a random noise.
Spectral uses Frobenius norm – should have better resilience to additive noise.
Due to the global optimal solution, Relative Entropy shows comparable results.
Zass & Shashua
Limitations of Graphs
Spectral MatchingLeordeanu05 withCour06
preprocessing
Spectral MatchingLeordeanu05
Our scheme (graphs)
Our scheme (hypergraphs,z=60)
Affine Transformation (doesn’t preserve distances)
random distortion additional points to a single graph to both
graphs
Zass & Shashua
Feature Matching inComputer Vision
Based solely on local appearance: Different features might look the same. Same feature might look differently.
Describe objects by local features (e.g., SIFT).
Match objects by matching features.
Zass & Shashua
Global Affine Transformation
Images from: www.robots.ox.ac.uk/vgg/research/affine/index.html
Spectral Graph Matching Hypergraph Matching based on distances based on area ratio
10/33 mismatches no mismatches
Zass & Shashua
Non-rigid Matching Match first and last frames of a 200
frames video (6 seconds), using [Torresani & Bregler, Space-Time Tracking, 2002]
features.
Videos and points from: movement.stanford.edu/nonrig/
Zass & Shashua
Non-rigid Matching
Videos and points from: movement.stanford.edu/nonrig/
Zass & Shashua
Summary Structure translates to hypergraphs,
not graphs. Probabilistic interpretation leads to a simple
connection between input and output:
Globally Optimal solution underRelative Entropy (Maximum Likelihood).
Efficient for both graphs and hypergraphs. Apply to graph matching problems as well.
XS d