Post on 10-Jul-2020
Computational Models of Perceptual Organization
Stella X. Yu
Robotics Institute
Carnegie Mellon University
Center for the Neural Basis of Cognition
Yu: PhD Thesis 2003 – p.1/60
What Is Perceptual Organization
��
�
�
�
�
�
�
��
��
�
�
�
� �
�
� �
�
�
�
�
���
�
�
�
��
�
�
�
��
�
�
�
��
�
�
� �
�
�
�� ��
�
�
��
�
�
�
�
�
��
�
�
�
�
�
� �
�
��
��
��
�
�
�
�
��
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
��
�
�
�
�
�
���
�
�
� �
�
� �
�
�
�
�
��
��
�
�
�
��
�� �
��
��
��
��
��
�
�
�
�
�
�
� �
��
�
�
��
�
�
�
��
�
�
�
�
���
�
�
�
�
�
�
�
�
�
�
��
�
�
��
�
�
�
�
�
�
�
� ��
�
�
�
�
�
�
�
�
� �
��
��
�
��
�
�
�
�
�
� �
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
��
�
�
� �
�
Yu: PhD Thesis 2003 – p.2/60
What Is Perceptual Organization
�
�
��
�
�
�
�
�
��
��
�
�
�
�
�
�
�
�
��
�
�
�
�
��
�
��
�
�
�
�
� �
��
�
�
�� �
�
�
�
�
�
�
�
��
�
�
�
�
�
� ��
�
�
�
�
�
��
�
�
�
�
�
�
�
�
�
�
�
��
�
��
�
�
�
�
�
�
�
�
�
�
���
��
�
�
��
�
�
�
�
�
�
�
�
�
�
�
�
��
�
�
�
�
�
�
�
�
�
�
�
�
�
��
�
��
�
��
�
�
�
�
��
��
�
�
�
�
�
�
�
��
�
�
�
��
�
�
�
�
�
�
�
�
�
��
�
�
�
�
�
�
�
�
�
�
�
��
�
�
�
�
�
�
�
�
��
�
�
��
�
�
��
�
�
�
�
�
��
�
�
�
�
�
��
��
��
�
�
�
�
�
��
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
Yu: PhD Thesis 2003 – p.2/60
What Is Perceptual Organization
��
�
�
�
�
�
�
��
��
�
�
�
� �
�
� �
�
�
�
�
���
�
�
�
��
�
�
�
��
�
�
�
��
�
�
� �
�
�
�� ��
�
�
��
�
�
�
�
�
��
�
�
�
�
�
� �
�
��
��
��
�
�
�
�
��
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
��
�
�
�
�
�
���
�
�
� �
�
� �
�
�
�
�
��
��
�
�
�
��
�� �
��
��
��
��
��
�
�
�
�
�
�
� �
��
�
�
��
�
�
�
��
�
�
�
�
���
�
�
�
�
�
�
�
�
�
�
��
�
�
��
�
�
�
�
�
�
�
� ��
�
�
�
�
�
�
�
�
� �
��
��
�
��
�
�
�
�
�
� �
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
��
�
�
� �
�
Yu: PhD Thesis 2003 – p.2/60
What Is Perceptual Organization
(Martin et al)
Yu: PhD Thesis 2003 – p.3/60
What Is Perceptual Organization
� multiple choices� a variety of features� content-dependent
•••
��
�
��
��
�
���
�
� ��
� �
� � �
��
��
���
�
�
����
�
�
��
�
� ��
�
�
�� �
��
�� � �
����
���
�
���
��
��
�� �
�
��
���
�
�
�
�
�
��
�
�
�
�
�
�� �
�
�
�
�
�
��
�
��
��
�
��
�
���
�
�� �
�
� �
� �
�
�
���
��
�� ��� � �
� ��
��
���
���
�
�
�
�
�
� �
��
�
�
��
�
�
�
��
�
��
�
�����
�
��
�
�
�
��
��
�
�
��
�
�
�
�
�
�
�� ��
�
�
�
�
��
�
�
� ��
� ��
�
��
�
�
�
�
�
� ��
�
�
�
�
�
��
�
�
�� �
��
�
�
��
�
���
�
� �
�
� one choice� single feature� content-free
Yu: PhD Thesis 2003 – p.4/60
Why Perceptual Organization
image
recognition
Yu: PhD Thesis 2003 – p.5/60
Why Perceptual Organization
15
9
Mahamud multi-object detector
Yu: PhD Thesis 2003 – p.6/60
Why Perceptual Organization
Schneiderman face detector
Yu: PhD Thesis 2003 – p.7/60
Traditional Use of Perceptual Organization
image
segmentation
figure-ground
recognitionperceptual organization
sequential processing (Marr, Lowe, Witkin, Tenenbaum, ...)
Yu: PhD Thesis 2003 – p.8/60
Perceptual Organization without Object Knowledge
difficult and brittle
(Canny, Geman & Geman, Shah & Mumford, Witkin, Jacobs, ...)
Yu: PhD Thesis 2003 – p.9/60
Our Overall Approach
perceptual organization
Pragnanz
recognition
grouping figure-ground
Yu: PhD Thesis 2003 – p.10/60
Our Overall Approach
perceptual organization
interactive processing (Grossberg, McClelland,Grenandar,Mumford,Lee,...)
Pragnanz
recognition
grouping figure-ground
Yu: PhD Thesis 2003 – p.10/60
Our Overall Approach
perceptual organization
interactive processing (Grossberg, McClelland,Grenandar,Mumford,Lee,...)
Pragnanz
recognition
grouping figure-ground
A criterion
A fast solution
A wide range of images
Yu: PhD Thesis 2003 – p.11/60
Outline1. Computational framework: spectral clustering
2. Expand the repertoire of grouping cues: dissimilarity
3. Guide grouping with partial cues
4. Guide grouping with object knowledge
5. Summary and future work
+
�
��
�
��
�
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
1 2 3 4Yu: PhD Thesis 2003 – p.12/60
Generative Approach for Data Clustering
��
�
�
�
�
�
�
��
��
�
�
�
� �
�
� �
�
�
�
����
�
�
�
��
�
�
�
��
�
�
�
��
�
�
� �
�
�
�� ��
�
�
��
�
�
�
�
�
��
�
�
�
�
�
� �
�
��
��
��
�
�
�
�
��
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
��
�
�
�
�
�
���
�
�
� ��
� �
�
�
�
�
��
��
�
�
�
���� �
��
��
��
��
���
�
�
�
�
�
� �
��
�
�
��
�
�
�
��
�
�
�
�
���
��
�
�
�
�
�
�
�
�
��
�
�
��
�
�
�
�
�
�
�
� ���
�
�
�
�
�
�
�
� �
�� �
�
�
��
�
�
�
�
�
� �
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
��
�
�
� �
�
Key: Assumptions on the global structure of the data
Pros: Intuitive interpretation; analysis = synthesis
Cons: Model inadequacy and computational intractability
Yu: PhD Thesis 2003 – p.13/60
Discriminative Approach for Clustering
��
�
�
�
�
�
�
��
��
�
�
�
� �
�
� �
�
�
�
����
�
�
�
��
�
�
�
��
�
�
�
��
�
�
� �
�
�
�� ��
�
�
��
�
�
�
�
�
��
�
�
�
�
�
� �
�
��
��
��
�
�
�
�
��
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
��
�
�
�
�
�
���
�
�
� ��
� �
�
�
�
�
��
��
�
�
�
���� �
��
��
��
��
���
�
�
�
�
�
� �
��
�
�
��
�
�
�
��
�
�
�
�
���
��
�
�
�
�
�
�
�
�
��
�
�
��
�
�
�
�
�
�
�
� ���
�
�
�
�
�
�
�
� �
�� �
�
�
��
�
�
�
�
�
� �
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
�
��
�
�
� �
�
0.70.1
Key: Same group or not
Pros: Adaptable to all data structures; tractable computation
Cons: No interpretation of the groups
Yu: PhD Thesis 2003 – p.14/60
Grouping in a Graph-Theoretic Framework
�
�
�
��
�
�
�
�
�
��
Yu: PhD Thesis 2003 – p.15/60
Grouping in a Graph-Theoretic Framework
�
�
�
��
�
�
�
�
�
��
Representation: G = {V, E,W} = { nodes, edges, weights }
Yu: PhD Thesis 2003 – p.15/60
Grouping in a Graph-Theoretic Framework
�
�
�
��
�
�
�
�
�
��
Representation: G = {V, E,W} = { nodes, edges, weights }
Clustering: ΓKV = {V1, . . . , VK} = K-way node partitioning
(Shi & Malik, Zabih, Boykov, Veksler, Kolmogorov,...)
Yu: PhD Thesis 2003 – p.15/60
Links in Graph Cuts
�
�
�
��
�
�
�
�
�
��
Yu: PhD Thesis 2003 – p.16/60
Links in Graph Cuts
�
�
�
��
�
�
�
�
�
��
�
�
�
��
�
�
�
�
�
��
P Q
Yu: PhD Thesis 2003 – p.16/60
Links in Graph Cuts
�
�
�
��
�
�
�
�
�
��
P Q
links(P, Q) =∑
p∈P, q∈Q
W (p, q)
Yu: PhD Thesis 2003 – p.16/60
Degree in Graph Cuts
�
�
�
��
�
�
�
�
�
��
P Q
degree(P) = links(P, V)
Yu: PhD Thesis 2003 – p.17/60
Linkratio in Graph Cuts
�
�
�
��
�
�
�
�
�
��
P Q
linkratio(P, Q) =links(P, Q)
degree(P)
Yu: PhD Thesis 2003 – p.18/60
Goodness of Grouping in Graph Cuts
�
�
�
��
�
�
�
�
�
��
P
V \ P
Maximize within-group connections: linkratio(P, P)
Minimize between-group connections: linkratio(P, V \ P)
Equivalent: linkratio(P, P) + linkratio(P, V \ P) = 1
Yu: PhD Thesis 2003 – p.19/60
K-Way Normalized Cuts
�
�
�
��
�
�
�
�
�
��
V1 V2
V3
max knassoc(ΓKV ) =
1
K
K∑
l=1
linkratio(Vl, Vl)
min kncuts(ΓKV ) =
1
K
K∑
l=1
linkratio(Vl, V \ Vl)
Yu: PhD Thesis 2003 – p.20/60
A Principled Solution to Normalized Cuts
max knassoc(ΓKV ) =
1
K
K∑
l=1
linkratio(Vl, Vl)
NP complete even for K = 2 and planar graphs
Fast solution to find near-global optima:
1. Find global optima in the relaxed continuous domainoptima = eigenvectors of (W,D) × rotations
2. Find a discrete solution closest to continuous optimacloseness = measured in L2 norm between solutions
Yu: PhD Thesis 2003 – p.21/60
A Principled Solution to Normalized Cuts
max knassoc(ΓKV ) =
1
K
K∑
l=1
linkratio(Vl, Vl)
NP complete even for K = 2 and planar graphs
Fast solution to find near-global optima:
1. Find global optima in the relaxed continuous domainoptima = eigenvectors of (W,D) × rotations
2. Find a discrete solution closest to continuous optimacloseness = measured in L2 norm between solutions
Yu: PhD Thesis 2003 – p.21/60
A Principled Solution to Normalized Cuts
max knassoc(ΓKV ) =
1
K
K∑
l=1
linkratio(Vl, Vl)
NP complete even for K = 2 and planar graphs
Fast solution to find near-global optima:
1. Find global optima in the relaxed continuous domainoptima = eigenvectors of (W,D) × rotations
2. Find a discrete solution closest to continuous optimacloseness = measured in L2 norm between solutions
Yu: PhD Thesis 2003 – p.21/60
A Principled Solution to Normalized Cuts
max knassoc(ΓKV ) =
1
K
K∑
l=1
linkratio(Vl, Vl)
NP complete even for K = 2 and planar graphs
Fast solution to find near-global optima:
1. Find global optima in the relaxed continuous domainoptima = eigenvectors of (W,D) × rotations
2. Find a discrete solution closest to continuous optimacloseness = measured in L2 norm between solutions
Yu: PhD Thesis 2003 – p.21/60
A Principled Solution to Normalized Cuts
max knassoc(ΓKV ) =
1
K
K∑
l=1
linkratio(Vl, Vl)
NP complete even for K = 2 and planar graphs
Fast solution to find near-global optima:
1. Find global optima in the relaxed continuous domainoptima = eigenvectors of (W,D) × rotations
2. Find a discrete solution closest to continuous optimacloseness = measured in L2 norm between solutions
Yu: PhD Thesis 2003 – p.21/60
Step 1: Find Continuous Global Optima
+O
Yu: PhD Thesis 2003 – p.22/60
Step 1: Find Continuous Global Optima
continuous optima
+O
�
X∗
eigensolve
Yu: PhD Thesis 2003 – p.22/60
Step 2: Discretize Continuous Optima
+
�
�
eigensolve
O
X∗
X∗(0)
Yu: PhD Thesis 2003 – p.23/60
Step 2: Discretize Continuous Optima
+
�
�
�
eigensolve
initialize
O
X∗
X∗(0)
X∗(0)
Yu: PhD Thesis 2003 – p.23/60
Step 2: Discretize Continuous Optima
+
�
�
�
�eigensolve
initialize
refine
O
X∗
X∗(0)
X∗(0)
X∗(1)
Yu: PhD Thesis 2003 – p.23/60
Step 2: Discretize Continuous Optima
+
�
�
�
�
�eigensolve
initialize
refine
O
X∗
X∗(0)
X∗(0)
X∗(1)X∗(1)
Yu: PhD Thesis 2003 – p.23/60
Step 2: Discretize Continuous Optima
+
�
�
�
�
�
�
eigensolve
initialize
refine
O
X∗
X∗(0)
X∗(0)
X∗(1)X∗(1)
X∗(2)
Yu: PhD Thesis 2003 – p.23/60
Step 2: Discretize Continuous Optima
+
�
�
��
�
�
�
eigensolve
initialize
refine
O
X∗
X∗(0)
X∗(0)
X∗(1)X∗(1)
X∗(2)X∗(2)
Yu: PhD Thesis 2003 – p.23/60
Step 2: Discretize Continuous Optima
+
�
�
��
�
�
�
eigensolve
initialize
refine
converge
O
X∗
X∗(0)
X∗(0)
X∗(1)X∗(1)
X∗(2)X∗(2)
Final solution: (X∗(2), X∗(2))
Yu: PhD Thesis 2003 – p.23/60
Pixel Similarity based on Intensity Edges
1
2
3
image oriented filter pairs edge magnitudes
Yu: PhD Thesis 2003 – p.24/60
Discrete Optima Generated by Eigenvectors
K = 4 : 0.9901 0.9899 0.9881
Not many local discrete optima, all good quality
Yu: PhD Thesis 2003 – p.25/60
Discrete Optima Generated by Eigenvectors
K = 4 : 0.9901 0.9899 0.9881
Not many local discrete optima, all good quality
Yu: PhD Thesis 2003 – p.25/60
Multiclass Real Image Segmentation
Yu: PhD Thesis 2003 – p.26/60
Outline1. Computational framework: spectral clustering
2. Expand the repertoire of grouping cues: dissimilarity
3. Guide grouping with partial cues
4. Guide grouping with object knowledge
5. Summary and future work
+
�
��
�
��
�
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
1 2 3 4Yu: PhD Thesis 2003 – p.27/60
Perceptual Popout
Yu: PhD Thesis 2003 – p.28/60
Perceptual Popout
Yu: PhD Thesis 2003 – p.28/60
Goodness of Grouping: Attraction and Repulsion
�
�
�
��
�
�
�
�
�
��
P
V \ P
Maximize within-group attraction: linkratio(P, P;A)
Minimize between-group attraction: linkratio(P, V \ P;A)
Equivalent: linkratio(P, P;A) + linkratio(P, V \ P;A) = 1
Yu: PhD Thesis 2003 – p.29/60
Goodness of Grouping: Attraction and Repulsion
�
�
�
��
�
�
�
�
�
��
P
V \ P
Maximize between-group repulsion: linkratio(P, V \ P;R)
Minimize within-group repulsion: linkratio(P, P;R)
Equivalent: linkratio(P, P;R) + linkratio(P, V \ P;R) = 1
Yu: PhD Thesis 2003 – p.29/60
Normalized Cuts with Attraction and Repulsion
• Criteria
knassoc(ΓKV ) =
1
K
K∑
l=1
links(Vl, Vl;A) + links(Vl, V \ Vl;R)
degree(Vl;A) + degree(Vl;R)
kncuts(ΓKV ) =
1
K
K∑
l=1
links(Vl, V \ Vl;A) + links(Vl, Vl;R)
degree(Vl;A) + degree(Vl;R)
• Equivalent weight matrix and degree matrix
W = A −R + DR
D = DA +DR
Yu: PhD Thesis 2003 – p.30/60
Normalized Cuts with Attraction and Repulsion
• Criteria
knassoc(ΓKV ) =
1
K
K∑
l=1
links(Vl, Vl;A) + links(Vl, V \ Vl;R)
degree(Vl;A) + degree(Vl;R)
kncuts(ΓKV ) =
1
K
K∑
l=1
links(Vl, V \ Vl;A) + links(Vl, Vl;R)
degree(Vl;A) + degree(Vl;R)
• Equivalent weight matrix and degree matrix
W = A −R + DR
D = DA +DR
Yu: PhD Thesis 2003 – p.30/60
Negative Weights and Regularization
• Negative weights:
W = A − R
= (positive entries + offset) − (negative entries + offset)
• Equivalent matrices: (W + Doffset, D + 2Doffset)
• Regularization: increase the degrees of nodes withoutchanging the sizes of weights between two nodes.
• Decrease the sensitivity of linkratio for nodes with littleconnections.
Yu: PhD Thesis 2003 – p.31/60
Negative Weights and Regularization
• Negative weights:
W = A − R
= (positive entries + offset) − (negative entries + offset)
• Equivalent matrices: (W + Doffset, D + 2Doffset)
• Regularization: increase the degrees of nodes withoutchanging the sizes of weights between two nodes.
• Decrease the sensitivity of linkratio for nodes with littleconnections.
Yu: PhD Thesis 2003 – p.31/60
Negative Weights and Regularization
• Negative weights:
W = A − R
= (positive entries + offset) − (negative entries + offset)
• Equivalent matrices: (W + Doffset, D + 2Doffset)
• Regularization: increase the degrees of nodes withoutchanging the sizes of weights between two nodes.
• Decrease the sensitivity of linkratio for nodes with littleconnections.
Yu: PhD Thesis 2003 – p.31/60
Negative Weights and Regularization
• Negative weights:
W = A − R
= (positive entries + offset) − (negative entries + offset)
• Equivalent matrices: (W + Doffset, D + 2Doffset)
• Regularization: increase the degrees of nodes withoutchanging the sizes of weights between two nodes.
• Decrease the sensitivity of linkratio for nodes with littleconnections.
Yu: PhD Thesis 2003 – p.31/60
Roles of Attraction, Repulsion, Regularization
attraction + repulsion + regularizationYu: PhD Thesis 2003 – p.32/60
Segmentation with Repulsion and Regularization
attraction attraction, repulsion and regularization
Yu: PhD Thesis 2003 – p.33/60
Segmentation with Repulsion and Regularization
attraction attraction, repulsion and regularization
Yu: PhD Thesis 2003 – p.34/60
Segmentation with Repulsion and Regularization
attraction attraction, repulsion and regularization
Yu: PhD Thesis 2003 – p.35/60
Outline1. Computational framework: spectral clustering
2. Expand the repertoire of grouping cues: dissimilarity
3. Guide grouping with partial cues
4. Guide grouping with object knowledge
5. Summary and future work
+
�
��
�
��
�
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
1 2 3 4Yu: PhD Thesis 2003 – p.36/60
Grouping with Partial Cues
+ ⇒
Yu: PhD Thesis 2003 – p.37/60
Basic Formulation: Grouping with Constraints
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
maximize ε(ΓKV )
Yu: PhD Thesis 2003 – p.38/60
Basic Formulation: Grouping with Constraints
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
maximize ε(ΓKV )
subject to X(i, l) = X(j, l)
Yu: PhD Thesis 2003 – p.38/60
Computing Constrained Normalized Cuts
• Constrained eigenvalue problem
• Efficient solution using a projector onto the feasible space
• Generalize Rayleigh-Ritz theorem to projected matrices
Yu: PhD Thesis 2003 – p.39/60
Why Simple Constraints Are Insufficient
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
natural grouping
Yu: PhD Thesis 2003 – p.40/60
Why Simple Constraints Are Insufficient
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
natural grouping
Yu: PhD Thesis 2003 – p.40/60
Why Simple Constraints Are Insufficient
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
natural grouping
Yu: PhD Thesis 2003 – p.40/60
Why Simple Constraints Are Insufficient
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
natural grouping constrained grouping
Yu: PhD Thesis 2003 – p.40/60
Why Simple Constraints Are Insufficient
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
natural grouping constrained grouping
Yu: PhD Thesis 2003 – p.40/60
Why Simple Constraints Are Insufficient
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
natural grouping constrained grouping
Yu: PhD Thesis 2003 – p.40/60
Remedy: Propagate Constraints
• General formulation:
maximize ε(ΓKV )
subject to S · X(i, l) = S · X(j, l)
• Normalized cuts:
0.8
0.2
0.40.3
0.3
S = P, or∑
k
PikX(k, l) =∑
k
PjkX(k, l), or(P T U)TX = 0
Yu: PhD Thesis 2003 – p.41/60
Remedy: Propagate Constraints
• General formulation:
maximize ε(ΓKV )
subject to S · X(i, l) = S · X(j, l)
• Normalized cuts:
0.8
0.2
0.40.3
0.3
S = P, or∑
k
PikX(k, l) =∑
k
PjkX(k, l), or(P T U)TX = 0
Yu: PhD Thesis 2003 – p.41/60
Clustering Points with Sparse Grouping Cues
��
��
��
��
��
��
��
�
��
�
��
��
��
��
��
��
��
�
��
��
��
��
�� �
�
��
��
��
��
��
��
��
�
�
��
��
� ��
�� � �
��
��
� ��
�� ��
�
��� �
�
�� �
��
��
� � ��
�� �
��
��
�
��
�
��
��
��
��
��
��
�� �
�
��
��
��
��
��
�
��
��
��
��
��
��
��
�
��
�
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
simple bias smoothed bias
Yu: PhD Thesis 2003 – p.42/60
Image Segmentation with Biased Grouping
Yu: PhD Thesis 2003 – p.43/60
Image Segmentation with Biased Grouping
no biasYu: PhD Thesis 2003 – p.43/60
Image Segmentation with Biased Grouping
no bias simple biasYu: PhD Thesis 2003 – p.43/60
Image Segmentation with Biased Grouping
no bias simple bias smoothed biasYu: PhD Thesis 2003 – p.43/60
Image Segmentation with Biased Grouping
Yu: PhD Thesis 2003 – p.44/60
Image Segmentation with Biased Grouping
Yu: PhD Thesis 2003 – p.44/60
Image Segmentation with Spatial Attention
Yu: PhD Thesis 2003 – p.45/60
Image Segmentation with Spatial Attention
Yu: PhD Thesis 2003 – p.45/60
Outline1. Computational framework: spectral clustering
2. Expand the repertoire of grouping cues: dissimilarity
3. Guide grouping with partial cues
4. Guide grouping with object knowledge
5. Summary and future work
+
�
��
�
��
�
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
1 2 3 4Yu: PhD Thesis 2003 – p.46/60
Object Segmentation
Yu: PhD Thesis 2003 – p.47/60
Our Approach to Object Segmentation
Yu: PhD Thesis 2003 – p.48/60
Our Approach to Object Segmentation
Yu: PhD Thesis 2003 – p.48/60
Our Approach to Object Segmentation
Yu: PhD Thesis 2003 – p.48/60
Our Approach to Object Segmentation
Yu: PhD Thesis 2003 – p.48/60
Our Approach to Object Segmentation
Yu: PhD Thesis 2003 – p.48/60
Our Approach to Object Segmentation
Yu: PhD Thesis 2003 – p.48/60
Joint Pixel-Patch Grouping: Criterion
A
B
C
ε(ΓKV ,ΓK
U ;A,B) =1
K
K∑
l=1
linkratio(Ul, Ul;B) · degree(Ul;B)
degree(Vl;A) + degree(Ul;B)
+1
K
K∑
l=1
linkratio(Vl, Vl;A) · degree(Vl;A)
degree(Vl;A) + degree(Ul;B)
Yu: PhD Thesis 2003 – p.49/60
Joint Pixel-Patch Grouping: Consistency
A
B
C
ΓKU = {U1, . . . , UK}, ΓK
V = {V1, . . . , VK}
Bias linking patches with their pixels
Yu: PhD Thesis 2003 – p.50/60
How Object Knowledge Helps Segmentation
pixel only pixel w/ ROI pixel-patch
541s 150s 110s
Yu: PhD Thesis 2003 – p.51/60
How Segmentation Helps Object Detection
image patch density segmentation
Yu: PhD Thesis 2003 – p.52/60
When Does Our Method Fail
image patch density segmentation
Yu: PhD Thesis 2003 – p.53/60
Equally Applicable to Multiple Objects
Yu: PhD Thesis 2003 – p.54/60
Contributions to Perceptual Organization
1. grouping and figure-ground in one framework
figure
ground
Yu: PhD Thesis 2003 – p.55/60
Contributions to Perceptual Organization
2. grouping integrated with spatial and object attention
Yu: PhD Thesis 2003 – p.56/60
Contributions to Graph Theory
1. new grouping cues
attraction repulsion
regularization depth
Yu: PhD Thesis 2003 – p.57/60
Contributions to Graph Theory
2. new graph partitioning techniques
+
�
��
�
��
�
K-way cuts directed cuts
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
��
biased cuts joint cutsYu: PhD Thesis 2003 – p.58/60
Future Work
1. Automatic selection of the number of classes.
2. A model-based view on spectral clustering.
3. A criterion for comparing two segmentations.
4. Closing a feedback loop.
5. Object representation.
6. Scaling up.
Yu: PhD Thesis 2003 – p.59/60
Acknowledgements
• Jianbo Shi, Tai Sing Lee, Takeo KanadeShyjan Mahamud, David TolliverJing Xiao, Vandi Verma
• CMU HumanID LabCMU Hebert Lab
• ONR N00014-00-1-0915NSF IRI-9817496NSF LIS 9720350NSF CAREER 9984706NIH EY 08098
Yu: PhD Thesis 2003 – p.60/60