chapter 2, part 2 relaxation - ucla statisticssczhu/courses/ucla/stat_232b/handouts/... · 2013. 4....
TRANSCRIPT
1
Chapter 2, part 2 relaxation
Chapter 2 covers topics to study the typical problems, data structures, and algorithmsfor inferring hierarchical and flat models.
Part 1: Search on hierarchical Generative representationsHeuristic search algorithms on And-Or graphs
-– Best First Search, A*, Generalized Best First Search.
Part 2: Search on flat descriptive representations2.1 Relaxation algorithm on line drawing interpretations 2 2 Belief propagation on poly-trees
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
2.2 Belief propagation on poly-trees2.3 Dynamic programming on chains.
Relaxation in a flat graph
Common properties:1. A graph representation G=<V, E>.
G could be directed undirected such as chain tree DAG lattice etcG could be directed, undirected, such as chain, tree, DAG, lattice, etc.
2. hard constraints or soft “energy” preference between adjacent vertices.
candidatelabel xi
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
Examples: constraint-and-satisfaction, line drawings, graph partition/coloring, Turbo coding,image segmentation, scene labeling, …
site i
label xi
2
Ex 1: Symbolic interpretation of line drawings (Waltz, 1960s)
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
Labeling the lines with 4 symbols
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
3
Excluding accidental views
Accidental views cause alignments(lines or junctions) which are unstable.
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
Examples of impossible realizations of corner by 3 faced object junctions
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
4
Finding all allowable junctions by considering all combinations of cubes in 8 quadrants in 3D
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
This is to exhaust all possible cases, all other non-accidental junctionswill be equivalent to one of these junctions.
Finding all allowable junctions by considering all combinations of cubes in 8 quadrants in 3D
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
5
All allowable 2-way / 3-way junctions
18 valid junctions out of208 possible combinations
(208= 16 + 64 + 64 + 64)
Remarks: in general vision modeling, the space of parts/objects/scenes is combinatorial,but the true elements in real world is tiny.This leads to huge constraints as in linedrawing interpretation.
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
For an alternative way, one may manually labela number of representative block objects andthen find the valid junctions from the label. Thismay not be exhaustive but is quite effective.In vision, we find vocabulary/elements this wayAs we couldn’t possibly enumerate all cases.
Finding a valid interpretation through relaxation:constraints-and-satisfaction
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
6
Finding a valid interpretation through relaxation:constraints-and-satisfaction
Note that what we compute hereis still candidate labels at eachis still candidate labels at eachvertex. When you have multipleways to label an object, then we have to extract the solution.
Discussion:
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
marginal beliefs vs. joint solutions
More complex labeling scheme by Waltz
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
7
Ex 2: 3D line drawing interpretation (Marill 91, Leclerc-Fishler’92)
1, Real 3D shapes2, Soft-energy
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
The energy function to optimize
SDA enforces equal angelsSDA stands for the standard deviation of
DP enforces co-planarityfor angels in each face.
SDA enforces equal angelsat each 3D vertex (thus toreach symmetry.
SDA stands for the standard deviation ofall angles at a vertex.
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
8
A typical result
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
More recent work on space reasoning on real images
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
Results from Yibiao Zhao, UCLA 2011.other reference: Lee, Gupta, Hebert and Kanade, Estimating spatial layout of rooms using
volumtric reasoning about objects and surfaces, NIPS 2010.
9
Ex 3: Belief propagation
b(x) = p(e- | x) p(x | e+) = (x) (x)
In a Markov chain representation
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
BP is equivalent to DP in a chain
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
10
on a flat graph (no-loop)
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
Message passing
Belief is the marginal posterior probability of a node x over all evidence e
b(x) = prob(x | e) = p(e- | x) p(x | e+) = (x) (x)
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
11
BP on Factor Graphs
, , , , , , , ,f x y z w v x y z z w z v
x
y
z
w
v
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
y
• The messages sent across the edges are the “local” marginal pdfs.
• The marginal pdf of z is the multiplication of all incoming messages
Belief Propagation Algorithm
x
y
z
w
v
~ \
z az a N z
m X m
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
\z f z
f N z
m m
12
• Nodes generate messages and send them to their neighbors
• BP provides the right answer if the graph has no loops
Belief Propagation
p g g p p
• Message propagation schedule is well defined
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
Loopy BP
• If graph has loops:– BP is no longer guaranteed to converge
– If it converges it is not necessarily to the right answer
– BP becomes an iterative algorithm where the message passing schedule is not well defined
• However, loopy BP has been shown to perform very well for many applications
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
13
Visiting order of nodes in the flat graph
The visiting order (scheduling) is an interesting problemThe visiting order (scheduling) is an interesting problem.
1, Discuss the case of global constraints and random belief updates
2, The visiting order in an And-Or graph can also be formulated in aBP algorithm.
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
Example: applying BP to channel coding
A binary-channel code can be seen as a constraint-and –satisfaction problem
(example from Dr. Andres I. Vila Casado EE, UCLA)
x1
x
x2 c11 0 1 0 1
1 1 1 1 0
c1
c2
x1 x2 x3 x4 x5
1 0 1 0 1
1 1 1 1 0
c1
c2
x1 x2 x3 x4 x5
1 0 1 0 1
1 1 1 1 0
c1
c2
x1 x2 x3 x4 x5
1 0 1 0 1
1 1 1 1 0
c1
c2
x1 x2 x3 x4 x5
1 0 1 0 1
1 1 1 1 0
c1
c2
x1 x2 x3 x4 x5
1 0 1 0 1
1 1 1 1 0
c1
c2
x1 x2 x3 x4 x5
1 0 1 0 1
1 1 1 1 0
c1
c2
x1 x2 x3 x4 x5
1 0 1 0 1
1 1 1 1 0
c1
c2
x1 x2 x3 x4 x5
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
x4
x5
x3
c3
c2
0 1 1 1 1c3
c2
0 1 1 1 1c3
c2
0 1 1 1 1c3
c2
0 1 1 1 1c3
c2
0 1 1 1 1c3
c2
0 1 1 1 1c3
c2
0 1 1 1 1c3
c2
0 1 1 1 1c3
c2
Each row is a constraint
14
Loopy BP for Channel decoding
Global function is the joint-probability of the received signal y and the true codeword x
1 1 1 1 1,..., , ,..., ,..., ,..., ,...,n n n n np x x y y p x x p y y x x
n
n
1 1 11
,..., , ,..., ,...,n n n i ii
p x x y y x x p y x
C 1 11
,..., ,...,n
n n i ii
p y y x x p y x
x1
x3
x2 c1
c
y1
y3
y2
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
x4
x5
3
c3
c2
y4
y5
y3
(Vila Casado)
Message-passing scheduling problem
How to schedule the message-passing algorithm in decoding?Pre-determined scheduling:
FloodingStandard Sequential Scheduling
Informed Dynamic Scheduling (IDS): Use the current state of messages in the graph to find
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
g g pthe next message to be propagated
(Vila Casado)
15
Simultaneous (Flooding) Schedule
On every iteration– All variable nodes are simultaneously updated– All check nodes are simultaneously updated
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
(Vila Casado)
Standard Sequential Schedule (SSS)
Update messages sequentially:
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
(Vila Casado)
16
100
Eb/No = 1.75 dB
Simultaneous (Flooding)LBP
Speed comparison
10-3
10-2
10-1
FE
R
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
31
0 5 10 15 20 25 30 35 40 45 5010
-4
Iterations
Andres I. Vila Casado, UCLA
Historical Milestones
Line drawing labeling (Waltz, 1972)
Relaxation-Labeling algorithm (Rosenfeld, Hummel and Zucker, 1976)
Belief propagation (Pearl, 1982)
Gibbs sampler (Geman and Geman, 1984)
Cluster sampling (Swendsen-Wang,1987)
For special cases: On Chains:
D i i
Stat 232B Statistical Computing and Inference in Vision S.C. Zhu
Dynamic programming (Bellman 1957)
= Belief propagation = Exact sampling (Gibbs sampler)