Visualization and Perception Laboratory Visualization and Perception Laboratory
Department of Computer Science and Engg., IIT Madras, Chennai – 600036, India.
//www.cs.iitm.ernet.in/~sdas/vp_lab/home.htm
//vplab.cs.iitm.ernet.in
Images, scenes, pictures
Model parameters, Object/Scene
representation Vision
Visualization
VISUAL VISUAL
PERCEPTION PERCEPTION
Flowchart illustrating the concept of visual perception.
Goal: Design of new computational frameworks, models and algorithms based on human perception and visualization to solve real-life complex problems in object/scene understanding and visual realism/visualization.
Two main categories of Areas/Applications
Object/Scene understanding
• BIOMETRY:
- multi-modal
- 3-D face and video
• Defense surveillance & GIS
• Generic 3-D Object Recognition
• Shape & Structure from 3D texture
• Foreground object detection
• CSS/MCC for MPEG-VII shape descriptors
Visualization
• Super-resolution texture synthesis
• Video based rendering
• Augmented Reality
• Modeling nature
Feature Extractor Probe
(Testing sample)
Identity or Accept/
Reject
Face Recognition Problem Face Recognition Problem
Face recognition can be defined as the identification of individuals
from images of their faces by using a stored database of faces labeled with people’s identities.
Matching 1:N or 1:1
Gallery (Training samples)
1-D Processing Input
Image
Edginess Image
Eigen Analysis
1-D Processing Input
Image
Edginess Image
Eigen Analysis Transformation
Function Eigenedginess:
Transformed edginess:
Face Recognition using PCA
Grey level image
Edge image
Edginess image
Results of EPCA
Artificial Neural Networks Lab IIT Madras
R e p r e s e n ta tio n P e r fo r m a n c e
E ig e n fa c e 1 4
E ig e n e d g e 2 4
E ig e n h ill 2 1
E ig e n e d g in e s s 5 6
Face Recognition performance
(Out of 80 faces)
Artificial Neural Networks Lab IIT Madras
Recognition performance due to variations in facial expression
Category % Eigenface 94 Eigenedginess 93 Eigenhill 77 Eigenedge 47
Artificial Neural Networks Lab IIT Madras
Modular Eigen-analysis or MPCA
VP Lab, CSE, IITM
Modular PCA
ea eb ec
+ E1
E2
E3
E4
w1 w2
w3
Weighted
Results of WMPCA
VP Lab, CSE, IITM
Results of WMPCA
VP Lab, CSE, IITM
Experimental Results of
VPLAB-FACE-RECOGNIZER
We have compared the performance of our face recognition method (patent application being filed), using four linear subspace methods (PCA, LDA – 1D & 2D) on three (3) Face Databases.
Near-frontal (pose variations)
10 40
ORL
Frontal, only illumination variation
42 60
PIE
Frontal, illumination and expression variation
11 15
YALE
Description No. of
samples /class No. of
subjects
Database
PEAK RECOGNITION ACCURACY (PRA) AND EQUAL ERROR RATE (EER) OF SUBBAND FACE REPRESENTATION INTEGRATED WITH PCA, LDA,
2D-PCA, 2D-LDA AND DCV WITH SUBJECT-SPECIFIC SUBBANDS OBTAINED, USING FOUR CRITERIA, ON YALE DATABASE
Experimental Results
Training images
RESULTS ON Yale DATABASE
Experimental Results
RESULTS ON PIE DATABASE
Training images
Experimental Results
Results on ORL DATABASE
Training images
Dual space Face Recognition using Feature Fusion
We propose a new face recognition technique by combining information from null and range space of within-class scatter of face space.
A new space named as dual space is constructed by merging two different set of discriminatory directions obtained from null and range space separately.
1) Covariance Sum and
2) Gramm-Schmidt Orthonormalization
We employ forward and backward selection techniques to
select the best set of discriminatory features from dual space.
Experimental Results
PERFORMANCE OF DUAL SPACE FACE RECOGNITION ON YALE, ORL AND PIE DATABASES.
PERFORMANCE OF NULL and RANGE SPACE FACE RECOGNITION ON YALE, ORL AND PIE DATABASES.
87.35 87.28
88.40 88.40
PIE
85.99 85.83
85.93 85.83
ORL
85.00 85.00
85.00 85.00
Yale
FS BS
FS BS
Databases
Gram-Schmidt Covariance Sum
73.14 86.35
PIE
71.67 79.17
ORL
75.00 81.67
Yale
Range Space Null Space
Methods
Dual space Face Recognition using Decision Fusion
Two classifiers are constructed based on optimal discriminatory directions obtained from null and range spaces separately.
Then these two classifiers are combined using Sum, Product and our proposed method of decision fusion.
Given a sample x, the output response from null and range space based classifiers are:
)]
( ),...,
( ),..,
( [
)
( x d 1 x d x d x
D Null = Null Null i Null C
)]
( ),...,
( ),..,
( [
)
( x d 1 x d x d x
D Range = Range Range i Range C
These vectors, called as response vectors, are the soft class labels provided by a classifier
Our aim here is to learn each classifier separately using
training information at decision level.
Proposed method of decision fusion
We use the training information (response vectors on a validation set) as feature vectors to construct LDA and
nonparametric LDA based eigenmodel.
Now new test response vectors are calculated in the eigenmodel and are used to replace the old ones.
Replace DP(x) by,
Responses of the classifier on a validation set of the database (which is disjoint from training and testing set) can be used as training information.
Sum Rule: Computes the soft class label vectors using,
Product Rule: Computes the soft class label vectors using,
Experimental Results
PERFORMANCE OF DUAL SPACE BASED FACE RECOGNITION ON YALE, ORL AND PIE DATABASES.
PERFORMANCE OF NULL AND RANGE SPACE FACE RECOGNITION ON YALE, ORL AND PIE DATABASES.
73.14 86.35
PIE
71.67 79.17
ORL
75.00 81.67
Yale
Range Space Null Space
Methods
100.00 100.00
83.85 83.72
PIE
97.50 96.67
81.67 81.67
ORL
86.67 86.67
86.67 86.67
Yale
P NLDA P LDA
Product Sum
Methods
Results of
Multi-modal Biometry
VP Lab, CSE, IITM
Decision Fusion of Face and Fingerprint
We have chosen face and fingerprint for their
universality and uniqueness properties respectively.
Feature level fusion is not possible in our case because of incompatible feature sets (e.g., fingerprint minutiae
and eigen-face coefficients).
The philosophy behind Decision Fusion:
¾ One of the classifiers can be best in terms of performance
¾ but the sets of patterns misclassified by different classifiers would not necessarily overlap.
This suggests that different classifiers offers
complementary information about the pattern to be
classified, for the same pattern classification problem .
Experimental Results
Assuming that face and fingerprint are statistically independent for an individual, we have associated an
individual from face database with an individual from fingerprint database to create a virtual subject.
Face Databases Used: PIE and ORL.
Fingerprint Databases Used: FP A and FP B (from NIST).
So get four possible combinations: (i) PIE & FP A , (ii) PIE
& FP B , (iii) ORL & FP A and (iv) ORL & FP B .
For face Databases we have taken two test cases and for
fingerprint four test cases.
Experimental Results
Base classifier's performance (in Percentage Accuracy) for four face and fingerprint databases (two each)
Base classifier's performance (in Percentage Accuracy) using
LDA for the face and fingerprint databases (two each)
Base classifier's performance (in Percentage Accuracy)
using nonparametric LDA for the face and fingerprint
databases (two each)
Experimental Results
Combined performance with PIE and FP A databases for
different Decision Fusion Strategies
Experimental Results
Combined performance with PIE and FP B databases for
different Decision Fusion Strategies
Experimental Results
Combined performance with ORL and FP A databases
for different Decision Fusion Strategies
Experimental Results
Combined performance with ORL and FP B databases
for different Decision Fusion Strategies
Overview of the approach for Road Overview of the approach for Road
detection from satellite images detection from satellite images
Satellite image is processed in four stages:
• First stage: Edge map is obtained from satellite image using 1-D processing
• Second stage: Edge map is post-processed using a set of binary morphological operations
• Third stage: Linear and curvilinear segments are
extracted using a GMM based extractor from the edge map
• Final stage: A FFNN is used to detect road segments from GMM based extractor output
Artificial Neural Networks Lab IIT Madras
Satellite image of a part of Chennai,
Source: NRSA, 1998. Same from Google-Earth, Nov-2005.
Satellite image of a part of Chennai (i) Post-processed edge map (ii) GMM output
Manually plotted road network using a GUI tool (iii) ANN output
Scanned map of same part
Satellite Image of a part of Chennai city.
Res: 5.8 m
Extracted road segments
using a hybrid approach
SATELLITE IMAGE – ROAD DETECTION
A satellite image
ANN output Road boundaries from ANN output
Output using GMM Manually plotted road network
Multi-spectral Satellite image, Res: 1m
Extracted road segments using
a hybrid approach
2 2 - - D TEXTURE SEGMENTATION D TEXTURE SEGMENTATION
Flow chart of the method of texture
classification
Filtering
Non-linearity
Smoothing
Normalizing non-linearity
Classifier Filter responses
Local energy function
Local energy estimates
Feature vectors
Input image
Segmented map
Image with five Texture regions
Segmentation result using:
Gabor + DWT, FCM
VP Lab, CSE, IITM
Results (Cont.)
I1 I2 I3 I4 I5
Segmented maps (DWT+DCT)
Input images
Results (Cont.)
I6 I7 I8 I9 I10
Input images
Segmented maps (DWT+DCT)
Results on SAR images
SAR Image
SAR Image
Fine segmented image
Fine segmented image
Coarse segmented image
Coarse segmented
image
Filtering
Smoothing using 2-D asymmetric Gaussian filter
Self-Organizing feature Map (SOM)
Smoothing using 2-D symmetric Gaussian filter
Edge detection using Canny operator
Stages of processing for texture edge detection.
Edge map
Edge Linking
Input image Input image Edge map
Experimental results of texture edge detection.
Texture Edge Detection
Results
Input image Edge map Input image Edge map Input image Edge map
Integrating Region and Edge
Information for Texture Segmentation
We have used a modified constraint satisfaction neural networks termed as Constraint Satisfaction Neural Network for Complementary Information Integration (CSNN-CII), which integrates the region and edge based approaches.
+
Input Image Segmented map before integration (Ref: [Rao 2004]) Edge map before
integration
(Ref: [Gupta 2006]) Segmented map
and Edge map after integration
Results
Input Image
Segmented map before integration
Edge map before integration
Segmented map and Edge map after integration
Results
Results
Input Image
Segmented map before integration
Edge map before integration
Segmented map
and Edge map
after integration
Results (Cont.)
Input Image
Segmented map before integration (Ref: [Shivani 04])
Segmented map and Edge map after integration Segmented map
after integration as
given in [Munoz 02]
Proposed Hierarchical Landform Classification
Input Image
Coastal Fluvial
DP I
DP II Dunes
Fuse Landform Segments Parabolic
Longitudinal Barchanoids Sandy plain
Inselberg Salt Flats
Desertic
Fuse Landform Segments
Active Channel
FP I
Flood Plains
Bars Ox-Bow
FP II FP III
Fuse Landform Segments
Class-Labeled Output Image
Class-Labeled Output Image
Alluvial Plains CP I
Creek Coastal Bars
Swamp
Forested Swamp
Sea Sandy
Beach
Alluvial plain
CP II CP III
Class-Labeled Output Image
Integrated Landform Processing
Landform Processing
Integrated Results for Coastal landform
Coastal Image (size: 1060 x 1062)
Final Output
Fused Output Image
Input Image
Scene Classification
Test image Segmentation using DWT features
Feature detection from segments
Modified PNN Output
class
Training samples
Classification of indoor vs. outdoor scenes using texture, color and shape features and probabilistic neural networks
Indoor image
Outdoor
image
Examples of correctly classified
indoor images
Examples of correctly classified
outdoor images
Examples of indoor images misclassified
as outdoor images
Examples of outdoor images misclassified
as indoor images
Results
83.1 89.6
Color + Texture + Shape
90.8 94.0
Color + Texture
83.8 90.4
Shape + Texture
71.5 89.2
Shape + Color
86.9 94.0
Texture
53.5 94.0
Color
66.5 63.5
Shape
Outdoor Indoor
Feature set
INDOOR VS. OUTDOOR
CLASSIFICATION ACCURACY (%) ON IITM-SCID WITHOUT INTEGRATING REGION AND EDGE INFROMATION
83.8 89.7
Color + Texture + Shape
92.3 94.3
Color + Texture
84.8 90.5
Shape + Texture
- -
-
- -
-
- -
-
- -
-
Outdoor Indoor
Feature set
INDOOR VS. OUTDOOR
CLASSIFICATION ACCURACY (%) ON IITM-SCID BY INTEGRATING REGION AND EDGE INFROMATION
Ref: [lalit06b]
3D Texture Analysis 3D Texture Analysis
using DWT:
using DWT:
Orientation estimation Orientation estimation
and 3D texture segmentation
and 3D texture segmentation
system ordinate
- co Surface )
, , (
system ordinate
- co Image
) , (
system ordinate
- co World )
, , (
−
−
−
s s
s
y z
x v u
z y x
) Elevation(
,
Azimuth ( φ ) θ
Polar – azimuth representation
Advantages:
1. Separable Analysis along horizontal and vertical directions 2. Representation is complete
2 2 2
)]
tan(
) (
[
] ) (
[ ) sec
( α
α
i
i r
i
t F u
u F
u f
f −
= +
The Spatial Frequency The Spatial Frequency Distribution (SFD)
Distribution (SFD)
Algorithm DWT
Dominant scale selection Scale to frequency
Select points
Voting
Scalogram Input image
Scalemap
SFD
Angle NFM
Regression
Results of Simulated Image set
Dyadic level-3 decomposition
0 10 20 30 40 45
Input images
Energy distribution Scalogram
Spectrum
High frequency image, Origin at center
Curve fitting using Least squares
Normalized Frequency Map (NFM)
Selection of Points on the NFM
Experimentally obtained SFD
Comparison of Experimentally obtained SFD and theoretical SFD
0 10 20 30 40 45
Experimentally obtained SFD Simulated SFD
computed errors
F=520 parabola exp sfd parabola exp sfd 0 -0.59695 4.678939 4.089852 4.505786 4.678939 4.089852 10 11.15202 13.93284 11.96059 2.746978 3.932839 2.108769 20 24.73764 22.60145 20.53271 4.737642 2.601446 0.532714 30 37.29204 33.17194 30.02636 7.292045 3.171942 0.481817 40 45.62462 42.56704 38.9246 5.624617 2.567039 1.075396
abs errors
0 1 2 3 4 5 6 7 8
1 2 3 4 5
angles
errors parabola
exp sfd
RESULTS Using dyadic Daubeschies
WAVELET FILTERS
4-channel
computed errors
parabola exp sfd parabola exp sfd
0 8.25 6.02 2.61 8.25 6.02 2.61
10 17.08 12.31 9.07 7.08 2.31 0.93
20 26.34 26.07 23.62 6.34 6.07 3.62
30 38.07 37.40 31.30 8.07 7.40 1.30
40 48.55 47.44 38.98 8.55 7.44 1.02
0 1 2 3 4 5 6 7 8 9
0 10 20 30 40
true angle
abs error
parabola exp sfd
RESULTS Using 4-CHANNEL
WAVELET FILTERS
Real world images (Super and Bovik)
Stadium Bleachers steps Steps rotated Venetian Blind Brick wall
Aluminum wall
-30 -20 -10 0
Grill vertical Grill horizontal Ventilator 0
oVentilator -20
oComparison of errors
0 4 8 12 16 20
br ick 55 _0 gr ill 30_
0 gr ill 0_1
0
sta di um _0 _7 0
ste ps _0 _7 0 Ve ne tia n B lind_
...
Ve nt ilo ato r_0 _20 Ven
tilo ato r_ 0_ 0
wa ll 0
wa ll 1 0
wa ll 2 0
wa ll 3 0
Input image
A b s o lu te E rr o r
Super& Bovik Dyadic
M-channel
Azimuth = 22.07 Elevation = 20 Azimuth=26 Elevation=-37
2.80 -22.80
Azimuth = -20
-0.98 22.98
Polar = 22
Error Computed
True angle
1.04 -38.04
Azimuth = -37
4.73 30.73
Polar = 26
Error Computed
True angle
Azimuth =30 Elevation=-54
-54.14 0.14 Azimuth = -54
1.66 28.34
Polar = 30
Error Computed
True angle
Real world images (Rebeiro & Hancock), in PAMI’01
Our lab. images
in 2006:
3-D Real world Texture image
Fuzzy segmentation using wavelet features
The problem of 3-D texture segmentation
M -c ha nn e l W a ve le t f ilt er ban k
1 2
M
Local VAR.
Compn.
Local POWER Compn.
Fe a tu re S et I Fea tu re Se t I I
Wavelet
Coefficients Post-processing
3-D Texture processing using
m-channel wavelet filterbank
3-D Real world Texture image
3-channel DWT
Fuzzy segmentation using wavelet features
Segmentation using 3-D (gradient) texture features.
Desired number of classes (input to FCM):
2 3
Image of
WATER_MIRROR
Segmentation results, with:
3-channel bi-orthogonal
filter
4-channel orthogonal
filter
Curved Texture surface
Directional wavelet-based
(4-channel) 3D texture features:
Future Work: Segmentation of 3D textures having a
combination of planar and curved objects.
GENERIC OBJECT GENERIC OBJECT
RECOGNITION RECOGNITION
(GOR)
(GOR)
What is Generic Object Recognition?
z Generic Class Recognition:
Recognition not restricted to a single class of objects, say only faces or cars.
Spans multiple classes having a wide variety of distinguishing features.
Aims at specifying to which subclass a previously unseen object belongs to, rather than an exact identification of the object.
Requires use of multimodal features, selecting and decision fusion of different classifiers.
Object Recognition can be generic in two ways:
z Object Recognition : Distinguishing between multiple objects within a class eg. face recognition, fingerprint recognition, vehicle recognition.
z Generic Object (multiclass) Recognition with Pose:
Pose-invariant object recognition for various class of objects.
Requires feature fusion and/or decision fusion to deal with multiple classes.
Use of 3D information (model for structure
representation) for pose detection and
visualization.
Proposed Framework
• Hierarchical framework with two stages of processing
Distance Transform based Feature
Extraction and Matching Select a set of rank-ordered
samples
…
Selected Samples
Average, Select sample with minimum distance Minimum distance Report Recognized
Object
Stop Test
Image
Linear Subspace Analysis-based Intelligent Choice
Image Gallery
A
D.T. based
…
Correlation Distances Eigendistances
…
Compute Eigenspace using 2D-PCA
Project
Training Samples onto eigenspace
Training Samples
Projected Samples
…
A
Samples from COIL Object Database Parametric Object Eigenspace using 1D PCA
z Manifold is constructed for object eigenspace by projecting object images onto object eigenspace and using cubic spline interpolation
[ ] ( ( , ) ( ) )
) ( )
( 2 )
( 1 )
(
. l p p , p ,..., k p T r p l p
r e e e x c
f = −
z Three most prominent eigenvectors shown. Dots correspond to projections of learning samples.
z Constant illumination. Thus, appearance is given by a curve with single parameter, rather than a surface.
Object Eigenspace Construction using 1D PCA
Generic Classifier for Intelligent Choice : 2D PCA
z Used to reduce search space to a small number of samples
z Eigenspace represents a compact model of object’s appearance
Parametric Eigenspaces for two objects
z Three most prominent eigenvectors shown. Dots correspond to projections of learning samples.
z Constant illumination. Thus, appearance is given by a
curve with single parameter, rather than a surface.
2D PCA (Cont.)
Eigenspace of two objects with distinct appearances
2D PCA (Cont.)
Eigenspace of two objects with similar appearances and
different shapes
Results for 2D PCA with Shape Matching
Results (First phase) of GOR
over 100 objects of COIL database
Percentage Accuracy of 2D PCA Percentage Accuracy of 2D PCA with Shape Matching
Comparison of percentage accuracies of 2D PCA and 2D PCA with
shape-based Matching for eigendimension =3
2D PCA Vs ICA
Comparisoin of Recognition Accuracies of 2D PCA and ICA on COIL-100 Database
88 90 92 94 96 98
Pose Increment for training
% Acccuracy
2D PCA 96.375 93.625 93.25 89.5 88.75
ICA 97 94.5 93.75 91.25 91
10 15 20 25 30
2D PCA EigenDimension = 10; No. of ICs = 110
Decision Fusion of 2D PCA and ICA
• Both 2D PCA and ICA contribute complementary information to the classification task.
• Classifier Combination
Test Image
…
Training Samples 2D PCA
ICA
Matching Score
Matching Score
Normalized Scores
Report Closest
Object
∑
Ref: [Kittler04]
Decision Fusion of 2D PCA and ICA
Performance of Combined Classifier using SUM Rule for Fusion on COIL-100 Database
80 85 90 95 100
Pose Increment for training
% Accuracy
2D PCA 96.375 93.625 93.25 89.5 88.75 84.125 82
ICA 97 94.5 93.75 91.25 91 86 84
2D PCA & ICA Fusion 97.375 94.625 93.75 91.375 91.375 86.875 84.25
10 15 20 25 30 35 40
2D PCA EigenDimension = 10; No. of ICs = 110
Shape Matching on Fused Generic Classifier
Comparision of effect of Shape Matching on 2D PCA, ICA and Combined Classifier
89 91 93 95 97 99
Pose Increment for training
% A c cu ra cy
2D PCA+Shape 97.5 93.875 93.75 91.375 90
ICA + Shape 97.25 95.5 93.75 90.625 90.5
Fusion + Shape 98.25 95.875 93.75 91.5 91.375
10 15 20 25 30
Comparison of Proposed Framework with present state of art techniques
*Results reported in literature
$
Our Implementation
96.03 1152
32
*SVM
97 400
96.639 3600 100
ICA
100 720 20
*2D PCA
97.694 98.25
95.468 96.375
% Accuracy 100
3600 400
3600 400
Test 720
Samples
100 100
$20
*No. of Objects
Proposed Fusion + Shape
Matching 1D PCA
Technique
Results:
Cluttered Scenes Extracted Object Using GrabCut
Recognized Object
from the database
Application: Recognition in Cluttered background
Cluttered Scenes Extracted Object Using GrabCut
Recognized Object
from the database
Application: Recognition in Cluttered background
Known Background Region
Known Foreground Region Unknown Region
Original Videos Segmented Moving object
Retrieved Model &
Rendered Image
Application: Recognition from Video
Application: Content-based Video Retrieval ( CBVR ):
Develop a system to retrieve the video similar to that of a given query.
Sample Videos in the database
Query Images
Retrieved videos from the database
Video no: 25 Video no: 1
1
10
25
4
33 Video
no
Use of Active Contour for texture object detection
Contour Initialization Detected Texture Boundary
Modeling of Texture Force
Modeling of Texture Force
Texture Feature Extraction Scalogram Estimation
Texture Feature Estimation
Segmentation Texture Image
Texture Feature Image
Contd…..
Modeling of Texture Force
Texture Feature Extraction Scalogram Estimation
Texture Feature Estimation
Segmentation Texture Image
Input Texture Image
Scalogram Estimation
Texture Force Modeling
Modeling of Texture Force
Texture Feature Estimation
Texture Feature Image
Texture Feature Image
Segmentation Output
Texture Feature
Image
Input Image
Experimental Results
Segmentation output
Texture Feature
Image
Input Image
Experimental Results Contd…..
Our proposed approach for segmentation of object with
hole, using the combination of Active Contour and GrabCut
SnakeCut
Integration of Active Contour and GrabCut
SnakeCut
Central rectangle is not part of the object
User selects object to be
segmented
Desired output
Sna keC ut
Problem in this Output:
Upper part of the object has been thrown off
Gr ab Cu t
Problem in this Output:
Hole is detected as part of the object
Active Contour
Real world Result-1
Here, objective is to crop the wheel from the input image
Cropped image should not contain any part of the
background
Snake Output
GrabCut Output
SnakeCut
Output
Real world Result-2
Here, objective is to crop the soldier from the input image
Cropped image should not contain any part of the
background
GrabCut Output
SnakeCut
Output
Snake Output
TEMPLATE MATCHING
TEMPLATE MATCHING
The Problem Definition
IMRN
IMT
Given a bitmap template (IMT) and a noisy bitmap image IMRN which contains IMT (believe me):
FIND OUT the location of IMT in IMRN !
Go to the next page for more:
Problem explanation for pessimists.
• IMRN (in previous page) is obtained by adding a large level of “Salt and Pepper” noise onto IMR bitmap image.
• IMT is also obtained from IMR as shown above.
IMT
IMR
The RESULT beats the
human EYE – comprehensively
IMR IMRN
IMT
Target Scene
Target
Scene
Face Detection
Face Detection
Super-resolution texture synthesis using Image Quilting
Minimum-error boundary cut.
Blocks constrained by overlap.
Random placement
of blocks.
Image Quilting
Input Image
Output for Patch Size of 32 Pixels and Overlap Size of 6 Pixels.
Output for Patch Size of 48 Pixels and Overlap Size of 9 Pixels.
Input Image
Output for Patch Size of 32 Pixels and Overlap Size of 6 Pixels.
Output for Patch Size of 48 Pixels and Overlap Size of 9 Pixels.
Image Quilting
Input Image
Output for Patch Size of 32 Pixels and Overlap Size of 6 Pixels.
Output for Patch Size of 48 Pixels and Overlap Size of 9 Pixels.
Input Image
Output for Patch Size of 32 Pixels and Overlap Size of 6 Pixels.
Output for Patch Size of 48 Pixels and Overlap Size of 9 Pixels.
Image Quilting
Input Image
Output for Patch Size of 32 Pixels and Overlap Size of 6 Pixels.
Output for Patch Size of 48 Pixels and Overlap Size of 9 Pixels.
Input Image
Output for Patch Size of 32 Pixels and Overlap Size of 6 Pixels.
Output for Patch Size of 48 Pixels and Overlap Size of 9 Pixels.
Image Quilting
Input Image
Output for Patch Size of 32 Pixels and Overlap Size of 6 Pixels.
Output for Patch Size of 48 Pixels and Overlap Size of 9 Pixels.
Input Image
Output for Patch Size of 32 Pixels and Overlap Size of 6 Pixels.
Output for Patch Size of 48 Pixels and Overlap Size of 9 Pixels.
Future Work (planned and unplanned) Future Work (planned and unplanned)
• Large scale visual Surveillance and monitoring
• 3D scene analysis
• Multi-sensor data fusion and super-resolution analysis
• Modeling of soft objects
• Mobile Reality
• Artificial life
• Visual Area Networking (VAN)
To be To be
Continued Continued
&
&
Thank You Thank You
June, 2007