Iterative Contraction and Merging Bayesian Sequential Partitioning JNDBSP 1 Manifold Learning Bosh Shih 2 O utline Introduction Principal Component Analysis PCA Linear Discriminant Analysis LDA ID: 796719
Download The PPT/PDF document "Outline Manifold Learning" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Outline
Manifold LearningIterative Contraction and MergingBayesian Sequential PartitioningJND-BSP
1
Slide2Manifold Learning
Bosh Shih2
Slide3Outline
IntroductionPrincipal Component Analysis (PCA)Linear Discriminant Analysis (LDA)Multi-Dimensional Scaling (MDS)Isometric Feature Mapping (
Isomap
)
Demo
3
Slide4Introduction
Question:Complicated Feature → High Dimension → Data (exponential growth)The curse of dimensionalityDifferent Appearance
Answer:
Dimensionality Reduction
Manifold Learning (one of them)
Visualize
Ref :
The Manifold Ways of Perception.
4
Slide5Introduction
Low dimensional manifold embedded in the high dimensional space
Ref :
中央
研究院週報
(
第
1058
期
)
流行學習與人臉辨識
5
Slide6Principal Component Analysis (PCA)
Ref :
Principal Component Analysis (PCA) ,
NTHU , Open Course Ware
6
Slide7Linear Discriminant Analysis (LDA)
↑
↓
PCA
LDA
Supervised Learning
U
nsupervised Learning
7
Slide8Multi-Dimensional Scaling (MDS)
Ref :
Modern Multidimensional Scaling:
Theory
and
Applications
古典
多維標度法
(MDS
) ,
線
代
啟示錄
Advanced
Introduction to Machine
Learning ,
CMU-10715
8
Slide9Isometric Feature Mapping (Isomap
) Ref : A Global Geometric Framework for Nonlinear Dimensionality Reduction.
Step1
:
Neighborhood Graph (
)
Step2
: S
hortest Paths (
)
Step3
:
9
Slide10Demo
Ref : http://www.math.ucla.edu/~wittman/mani Advanced Introduction to Machine Learning,
CMU-10715
10
Slide11Swiss Roll
11
Slide12Twin Peaks
12
Slide13Curvature
13
Slide14Corners
14
Slide1515
Slide16Toroidal Helix
16
Slide17Iterative Contraction and Merging
IEEEJia-Hao Syu , Sheng-Jyh Wang & Li-Chun Wang2017
Ref :
Hierarchical Image Segmentation based on
Iterative
Contraction and Merging
17
Slide18Introduction
Segmentation (ICM)a sequel of optimizationpixel-based
+
grid-based
region-based
hierarchical
multi-resolution
18
Slide19Outline
Pixel-basedAffinityEnergy FunctionOptimal SolutionJust-Noticeable Difference (JND)
Cells
Remnant Removal
19
Region-based
Color
Texture
Region Size
Spatial Intertwining
Slide20Pixel-based
Affinity
Energy Function
Ref :
Learning-based hierarchical graph for
unsupervised
matting and foreground estimation
color
mean
window
covariance
number
identity
regularization
20
Slide21Pixel-based
Optimal SolutionJust-Noticeable Difference (JND)
Smooth areas
identity
L
aplacian
21
Slide22Pixel-based
CellsRemnant Removal
22
Slide23Region-based
Color
23
Region-based
TextureGray-LevelWeber Local Descriptor (WLD)24
Ref :
WLD: A Robust Local Image Descriptor
Region-based
Region SizeSpatial Intertwining25
100 100 5
100
4 1.3
4 4 1
64
4 1.6
Slide26Region-based
ColorTextureRegion SizeSpatial Intertwining
26
Slide27ICM
27
Slide28Bayesian Sequential Partitioning
Journal of the American Statistical AssociationLuo Lu , Hui Jiang & Wing H. Wong2013
Ref : Multivariate Density Estimation by Bayesian Sequential Partitioning
28
Slide29Outline
Density EstimationKernel & HistogramBayesian Sequential Partitioning (BSP)
29
Slide30Density Estimation
Given : Data
Goal :
Estimate common density function
30
Slide31Kernel
Histogram
Kernel & Histogram
31
Ref :
What are the limitations of Kernel methods and when to use
kernel
?
http
://
en.wikipedia.org/wiki/Kernel_density_estimation
Slide32Importance Sampling
Monte Carlo SamplingImportance Sampling32
Ref
:
Sampling Methods
Importance
Sampling
Slide33Bayesian Sequential Partitioning (BSP)
CharacteristicNonparametricPiecewise Constant FunctionsBinary PartitionFast
33
Slide34Bayesian Sequential Partitioning (BSP)
Partition Score34
Ref :
Dirichlet
distribution
Slide35Bayesian Sequential Partitioning (BSP)
Sequential Importance Sampling
35
Bayesian Sequential Partitioning (BSP)
Sequential Importance SamplingImportance weightRandom Weights
36
2
1
1
Bayesian Sequential Partitioning (BSP)
Resampling37
Ref :
TLD (Tracking-Learning-Detection
)
Slide38Bayesian Sequential Partitioning (BSP)
Example : 38
Ref :
K.C. Lo ’s group-meeting PPT
Slide39Unsupervised Hierarchical Image Segmentation
based on Bayesian Sequential Partitioning IEEEHao-Wei Yeh, Chen-Yu Tseng, Tung-Yu Wu, Sheng-Jyh
Wang
2015
Ref :
Unsupervised Hierarchical Image Segmentation
based
on
Bayesian
Sequential Partitioning
39
Slide40Method
Split PhaseJust-Noticeable-Difference Bayesian Sequential Partitioning (JND-BSP
)
[1] color
differences among
pixels
[2]
just-noticeable
difference
prior
Merge PhaseHierarchical
40
Slide41Random Walk Model
41
JND-BSP
42
Slide43JND-BSP
43
Slide44Reference
The Manifold Ways of Perception.中央研究院週報 (第1058期) 流行學習與人臉辨識
Principal Component Analysis (PCA) , NTHU , Open Course
Ware
Modern Multidimensional Scaling: Theory and Applications
古典
多維標度法
(MDS) ,
線代啟示錄
Advanced
Introduction to Machine Learning ,
CMU-10715A Global Geometric Framework for Nonlinear Dimensionality Reduction.
http://www.math.ucla.edu/~wittman/maniAdvanced Introduction to Machine Learning, CMU-10715 Hierarchical Image Segmentation based on Iterative Contraction and
Merging
Learning-based hierarchical graph for unsupervised matting and foreground
estimation
Multivariate Density Estimation by Bayesian Sequential
Partitioning
What are the limitations of Kernel methods and when to use kernel
?
http://en.wikipedia.org/wiki/Kernel_density_estimationSampling Methods
Importance Sampling Dirichlet
distributionTLD (Tracking-Learning-Detection)K.C. Lo ’s group-meeting PPT
Unsupervised Hierarchical Image Segmentation based on Bayesian Sequential Partitioning
44
Slide4545