Pattern Recognition 1
Preface 4
Organization 6
Table of Contents 9
Motion and Tracking 9
A 3-Component Inverse Depth Parameterization for Particle Filter SLAM 14
Introduction 14
Notation 15
A Monocular PF-SLAM System 15
Inverse-Depth Parameterization and PF-SLAM 17
Experimental Results 18
Conclusion 23
References 23
An Efficient Linear Method for the Estimation of Ego-Motion from Optical Flow 24
Motivation 24
Model of Instantaneous Ego-Motion 25
Linear Method for Ego-Motion Estimation 25
Results 28
Discussion 31
Conclusion 32
References 32
Localised Mixture Models in Region-Based Tracking 34
Introduction 34
Foreground-Background Separation in Region-Based Pose Tracking 36
Localised Mixture Models 37
Case I: Static Background Image Available 38
Case II: Potentially Varying Background 39
Experiments 39
Summary 42
References 42
A Closed-Form Solution for Image Sequence Segmentation with Dynamical Shape Priors 44
Introduction 44
An Implicit Dynamic Shape Model 45
A Statistical Formulation of Sequence Segmentation 46
Experimental Results 51
Conclusion 53
References 53
Markerless 3D Face Tracking 54
Introduction 54
Surface Tracking 55
Algorithm 56
Implementation 58
Results 60
Discussion and Future Work 61
References 62
Pedestrian Recognition and Automotive Applications 9
The Stixel World - A Compact Medium Level Representation of the 3D-World 64
Introduction 64
Building the Stixel-World 66
Dense Stereo 66
Occupancy Grid 66
Free Space Computation 67
Height Segmentation 68
Stixel Extraction 69
Experimental Results 71
Future Work 71
Conclusion 72
References 73
Global Localization of Vehicles Using Local Pole Patterns 74
Introduction 74
Mobile Mapping Setup and Feature Extraction 75
Characteristics of Local Pole Patterns 76
A Local Pole Pattern Descriptor 77
The Curse of Dimensionality 77
Design of the Local Pattern Descriptor 79
Scalability 80
Conclusions 82
References 83
Single-Frame 3D Human Pose Recovery from Multiple Views 84
Introduction 84
Previous Work 85
Single-Frame 3D Pose Estimation 86
Overview 86
Shape Model 86
Pre-processing 86
Single-View Shape Detection 87
Multi-view Pose Verification 88
Gradient-Based Local Pose Optimization 89
Experiments 89
Conclusion and Further Work 92
References 93
Dense Stereo-Based ROI Generation for Pedestrian Detection 94
Introduction 94
Related Work 95
Dense Stereo-Based ROI Generation 96
Modeling of Non-planar Road Surface 96
Outlier Removal 97
System Integration 99
Experiments 99
Conclusions 102
References 102
Pedestrian Detection by Probabilistic Component Assembly 104
Introduction 104
ComponentDetection 106
Probabilistic Component Assembly 106
Probabilistic Appearance Model 107
Probabilistic Geometric Model 108
Experiments 109
Component Detection 109
Probabilistic Component Assembly 110
Conclusion 112
Future Work 112
References 112
High-Level Fusion of Depth and Intensity for Pedestrian Classification 114
Introduction 114
Related Work 115
Spatial Depth and Intensity Features 116
Fusion on Classifier-Level 118
Experiments 119
Experimental Setup 119
Experimental Results 120
Conclusion 122
References 122
Features 10
Fast and Accurate 3D Edge Detection for Surface Reconstruction 124
Introduction 124
Previous Work 125
The 3D Edge Detection Algorithm 125
Volume Function and 3D Edge Model 125
Phase 1: Voxel-Based Edge-Detection 126
Phase 2: Subpixel Refinement 127
Theoretical Analysis: Accuracy 128
Noisy 3D Images 129
Experiments: Accuracy and Speed 130
Conclusions 133
References 133
Boosting Shift-Invariant Features 134
Introduction 134
Shift-Invariant Features for Image Classification 135
Boosting Shift-Invariant Features 136
Boosting Smoothing Splines 136
Training Features Using a Boosting Framework 137
Experiments 139
USPS Handwritten Digit Recognition 139
UIUC Car Classification 140
Conclusion and Outlook 142
References 142
Harmonic Filters for Generic Feature Detection in 3D 144
Introduction 144
RelatedWork 145
Spherical Tensor Analysis 146
Preliminaries 146
Spherical Tensor Coupling 147
Spherical and Solid Harmonics 147
Spherical Derivatives 148
Harmonic Filters 148
Differential Formulation 149
The Voting Function 149
Pollen Porate Detection in Confocal Data 150
Reference Approaches 151
Training 151
Evaluation 151
Conclusion 152
References 153
Increasing the Dimension of Creativity in Rotation Invariant Feature Design Using 3D Tensorial Harmonics 154
Introduction 154
Preliminaries 155
Rotation Invariant Features from Tensorial Harmonics 157
Designing Features 157
Fast Computation of Tensorial Harmonic Coefficients 158
Transforming Cartesian Tensors into Spherical Tensors 159
Experiments 160
Conclusion 162
References 163
Training for Task Specific Keypoint Detection 164
Introduction 164
Related Work 165
Task Specific Keypoints 165
Detector Verification 165
Detector Learning 166
Training Samples 167
Keypoint Boosting 168
FuzzyWeak Learning by Domain-Partitioning 168
Weak Classifier 169
Detector Evaluation 169
Light and Seasonal Changes 170
Conclusions 172
References 172
Combined GKLT Feature Tracking and Reconstruction for Next Best View Planning 174
Introduction and Literature Review 174
Review of KLT and GKLT Tracking 175
KLT Tracking 176
Guided KLT Tracking 176
Combining GKLT Tracking and Robust 3D Reconstruction 177
Experimental Comparison of KLT, GKLT and GKLT3D Tracking 179
Comparison Using a Short Image Sequence 180
Comparison Using a Long Image Sequence 181
Comparison within an Information-Theoretic Approach for Next Best View Planning 182
Conclusion and Future Work 182
References 183
Single-View and 3D Reconstruction 10
Non-parametric Single View Reconstruction of Curved Objects Using Convex Optimization 184
Introduction 184
Existing Approaches to Single View Reconstruction 185
Contributions 185
Variational Framework for Single View Reconstruction 186
Variational Formulation 186
Silhouette Consistency 187
Volume Inflation 187
Optimization via Convex Relaxation 187
Interactive Single View Reconstruction 188
Image Segmentation 188
Interactive Editing 189
Implementational Issues 190
Experiments 191
Conclusion 193
References 193
Discontinuity-Adaptive Shape from Focus Using a Non-convex Prior 194
Introduction 194
Proposed Approach 196
Experiments 199
Conclusions 202
References 202
Making Shape from Shading Work for Real-World Images 204
Introduction 204
Our Three-Stage Approach 206
Finding the Region of Interest \u2013 Segmentation 206
Ensuring a Homogeneous Albedo \u2013 Inpainting by Edge Enhancing Diffusion 206
3-D Reconstruction \u2013 Shape from Shading 207
Real-World Experiments 208
Conclusions and Outlook 212
References 212
Learning and Classification 10
Deformation-Aware Log-Linear Models 214
Introduction 214
Image Distortion Model 215
Integrating the IDM into Log-Linear Models 215
Relationship to Gaussian Models 216
Maximum Approximation 217
Training Method 217
Pooling of Deformation Priors 218
Experimental Evaluation 218
Transfer to MNIST and Comparison to the State-of-the-Art 221
Conclusion 222
References 223
Multi-view Object Detection Based on Spatial Consistency in a Low Dimensional Space 224
Introduction 224
Object Model 226
Detection Algorithm 227
Evaluation 230
Parameter Settings 230
Experiments and Results 231
Conclusion 232
References 233
Active Structured Learning for High-Speed Object Detection 234
Introduction 234
Object Tracking in Image Sequences 235
Object Localization in Single Frames 235
Object Localization by Structured Regression 236
S-SVM Training by Delayed Constraint Generation 237
Fast Object Localization 238
Implementation with GPU Support 239
Training with Active Learning 239
Experimental Evaluation 241
Summary and Future Work 242
References 243
Face Reconstruction from Skull Shapes and Physical Attributes 245
Introduction 245
Background 246
Statistical Shape Models 246
Training Data 247
StatisticalModel Fitting 247
Method 248
Linear Skull Predictor 248
Attribute Prediction from Face Coefficients 250
Minimization and Face Prediction 250
Results 250
Skull and Face Prediction without Attributes 251
Attribute Prediction 253
Face Prediction 253
Conclusion 253
References 254
Sparse Bayesian Regression for Grouped Variables in Generalized Linear Models 255
Introduction 255
Sparse Bayesian-Grouped Regression Model 256
Application to Poisson and Binomial Models 259
Simulated Example 260
Application to MEMset Donor Dataset 261
Conclusion 263
References 263
Learning with Few Examples by Transferring Feature Relevance 265
Introduction 265
Related Work 266
Transfer of Feature Relevance (TFR) 267
Incorporation of TFR into Randomized Classifier Ensembles 268
Randomized Decision Trees 269
Estimating Feature Relevance Using Randomized Decision Forests 269
Experiments 270
Feature Extraction 271
Dirichlet Parameter 271
Influence of the Ensemble Size 272
TFR Improves Learning with Few Examples 272
Conclusion 273
FurtherWork 273
References 274
Pattern Recognition and Estimation 11
Simultaneous Estimation of Pose and Motion at Highly Dynamic Turn Maneuvers 275
Introduction 275
Object Model 276
Filter Approaches 278
Interacting Multiple Model (IMM) 278
Oracle Measurements and Adaptive Noise (EKF-OR) 278
Yaw Acceleration Model (EKF-YA) 279
Experimental Results 279
Filter Configuration 279
Synthetic Ground Truth 280
Real World Ground Truth 281
Filter Behaviour at Extreme Maneuvers 282
Conclusion 283
References 284
Making Archetypal Analysis Practical 285
Introduction 285
Archetypal Analysis 286
Properties of Archetypal Analysis 287
The Archetype Algorithm 288
Making Archetypal Analysis Practical 289
Focusing on a Working Set 289
Preselecting Archetypal Candidates 290
Application Examples 291
Summary 293
References 294
Fast Multiscale Operator Development for Hexagonal Images 295
Introduction 295
Hexagon Images 296
Hexagonal Operator Design 297
Operator Implementation 299
Fast Multiscale Operator Construction 300
Algorithmic Performance 301
Summary 303
References 303
Optimal Parameter Estimation with Homogeneous Entities and Arbitrary Constraints 305
Introduction 305
General Adjustment Model with Constraints 306
Mathematical Model 306
Objective Function and Estimation 307
Precision of the Estimates 308
Examples 309
Estimation of the Fundamental Matrix 309
Constrained Vanishing Points Determination 310
Conclusions 313
References 313
Detecting Hubs in Music Audio Based on Network Analysis 315
Introduction 315
Related work 316
Timbral Similarity Measures 316
Hubs and Hubness Measures 317
Hubness Score Based on Network Analysis 317
Evaluation Framework 318
Genre Classification 319
Collection Browsing 319
Clustering 319
Experimental Results 320
Methods and Data 320
Results 320
Conclusion 323
References 324
A Gradient Descent Approximation for Graph Cuts 325
Introduction 325
GraphCuts 327
Gradient Descent for Graph Cuts 328
Experiments 330
Conclusions 333
References 333
Stereo and Multi-view Reconstruction 11
A Stereo Depth Recovery Method Using Layered Representation of the Scene 335
Introduction 335
Method 337
Layered Representation of the Scene 337
Synchronous Energy Formulation for Stereo 339
Discontinuity Preserving Nonlinear Regularization 341
Experiments 342
Conclusions 343
References 344
Reconstruction of Sewer Shaft Profiles from Fisheye-Lens Camera Images 345
Introduction 345
Background 346
Problem Specification and Setting 346
3D Reconstruction Using Structure from Motion 347
Our Approach 348
Reconstruction of 3D Points 348
Geometric Correction 350
Shape Classification and Estimation 351
3D Model Creation from Cross-Sections 351
Experiments and Results 351
Evaluation with Real Labelled Data 351
Practical Issues: Robustness and Runtime 353
Resulting 3D Models 353
Conclusions 354
References 354
A Superresolution Framework for High-Accuracy Multiview Reconstruction 355
Introduction 355
Displacement Maps and Texture Superresolution 356
Variational Formulation 356
Transformation of the Data Term to the Surface 357
Solving for the Superresolution Texture 358
Solving for the Displacement Map 358
PDE-Based Energy Minimization on Texture Space 359
ConformalMaps and Differential Operators 360
Discretization and Computation Grid 361
Final Algorithm Implementation 363
Experiments 363
Conclusion 363
References 364
View Planning for 3D Reconstruction Using Time-of-Flight Camera Data 365
Introduction 365
Motivation 365
Literature Review 366
Paper Organization 367
Volumetric Planning Approach with TOF Camera Data 367
Evaluation \u2013 Assessing a Scan\u2019s Contribution 367
Multi-view Fusion 369
View Planning 370
Proposed TOF Camera Based View Planning Extension 371
Experimental Results 371
Conclusions and Future Work 373
References 374
Real Aperture Axial Stereo: Solving for Correspondences in Blur 375
Introduction 375
Scaling and Defocus in Axial Stereo 377
Structure and Focused Image Recovery 379
Experimental Results 380
Conclusions 383
References 383
Real-Time GPU-Based Voxel Carving with Systematic Occlusion Handling 385
Introduction 385
Related Work 386
GPU-Based Voxel Carving 387
Problem Description 387
Static Scene Occluders 388
GPU Implementation Using Cuda 389
Results 390
Evaluation on the Middlebury Dataset 390
Experiments in a Multi-camera Environment 392
Conclusion 393
References 393
Image-Based Lunar Surface Reconstruction 395
Introduction 395
Related Work 397
Algorithm 398
Normal Estimation 399
Credibility Map 400
Normal Integration 400
Results 402
Conclusion and Discussion 403
Image Analysis and Applications 12
Use of Coloured Tracers in Gas Flow Experiments for a Lagrangian Flow Analysis with Increased Tracer Density 405
Introduction 405
Experimental Setup 407
Reducing the Number of Ambiguities with Coloured Particles 408
Colour Classification 410
Results and Discussion 412
References 413
Reading from Scratch \u2013 A Vision-System for Reading Data on Micro-structured Surfaces 415
Introduction 415
Previous Works 416
Surface Reconstruction at Scratch Scale 417
Image Partitioning Using Spectral Analysis 418
Signal Tracking Using a Hidden Markov Model 419
Results 421
Perturbation Sensitivity of the Signal Tracking 421
Writing and Reading Data 421
Summary and Conclusion 423
References 424
Diffusion MRI Tractography of Crossing Fibers by Cone-Beam ODF Regularization 425
Introduction 425
Regularization of Diffusion MRI Datasets 426
Cone-Beam ODF Regularization (CB-REG) 427
Deterministic ODF-Based Fiber Tracking 429
Results 430
Conclusions 432
References 433
Feature Extraction Algorithm for Banknote Textures Based on Incomplete Shift Invariant Wavelet Packet Transform 435
Introduction 435
Shift Invariant Wavelet Packet Transform 436
Feature Extraction Algorithm for Shift Invariant Wavelet Packet Transform 438
Information Content and Stopping Criterion 438
Best Branch Algorithm 439
Experimental Results 441
Conclusions 443
References 444
Video Super Resolution Using Duality Based TV-L1 Optical Flow 445
Introduction 445
OpticalFlow 447
Super Resolution 449
Conclusion 453
References 454
HMM-Based Defect Localization in Wire Ropes \u2013 A New Approach to Unusual Subsequence Recognition 455
Introduction 455
Related Work 456
HMM-Based Anomaly Detection 457
Unusual Subsequence Detection 458
Experiments 460
Importance of the Amount of Training Data 461
Recovered Defect Area 461
Comparison to Time Invariant OCC 462
Conclusions 463
References 463
Beating the Quality of JPEG 2000 with Anisotropic Diffusion 465
Introduction 465
Diffusion-Based Image Compression 466
OurNovelCodec 467
Rectangular Subdivision 467
Improved Entropy Coding 468
Brightness Rescaling 469
Diffusivity Optimisation 469
Interpolation Swapping 469
File Format 470
Experiments 470
Conclusions 473
References 473
Decoding Color Structured Light Patterns with a Region Adjacency Graph 475
Introduction 475
Single-Shot Structured Light 476
Pattern Design 476
Watershed Segmentation 477
Region Adjacency Graph 478
Vertices 478
Edges 479
Pattern Decoding 480
Results 481
Conclusion and Future Work 483
References 483
Residual Images Remove Illumination Artifacts! 485
Introduction 485
Smoothing Operators 486
Residual Images Contain the Important Information 488
Removing Illumination Artifacts with Residual Images 490
Conclusions and Future Research 493
References 494
Superresolution and Denoising of 3D Fluid Flow Estimates 495
Introduction 495
Constrained Fluid Flow Denoising in 3D 496
Notation, Definitions 496
Solenoidal Projection 497
Lowpass Filtering 497
Vorticity Rectification 497
Velocity Restoration 498
Superresolution Approach 499
NumericalExperiments 499
Data Sets and Set-Up 499
Results and Comparison 501
Conclusion and Further Work 503
References 503
Spatial Statistics for Tumor Cell Counting and Classification 505
Introduction 505
The Proposed Method 506
Point Process Prior 507
Pointwise Likelihood 508
Two Algorithms for Fitting the Model to Images 508
Experimental Evaluation 509
Evaluation on SIMCEP Benchmark 511
Proliferation Fraction in Meningioma Cells 512
Discussion 513
References 513
Segmentation 13
Quantitative Assessment of Image Segmentation Quality by Random Walk Relaxation Times 515
Introduction and Related Work 515
Segment Homogeneity 516
Experiments 519
Conclusion 524
References 524
Applying Recursive EM to Scene Segmentation 525
Introduction 525
Problem Formulation 526
Segmentation Filter 527
E-step 528
M-step 529
S-Step 531
Experiments 532
Conclusion 534
References 534
Adaptive Foreground/Background Segmentation Using Multiview Silhouette Fusion 535
Introduction 535
Contribution 536
Learning 538
Learning the Background 538
Dealing with Illumination Influences 539
Learning the Foreground 540
3d Reconstruction by Probabilistic Fusion 540
Evaluation 541
Discussion 543
References 544
Evaluation of Structure Recognition Using Labelled Facade Images 545
Introduction 545
Facade Grammar 546
Reconstruction Process 547
MDL 548
Reconstruction of eTRIMS Facade Images 550
Conclusion 553
References 554
Using Lateral Coupled Snakes for Modeling the Contours of Worms 555
Introduction 555
The Model 557
Discrete Formulation 558
Forces on the Ends and Their Neighbors 559
Evolution and Multiscale Approach 560
LCS Model for Object Segmentation 560
Experiments 561
Conclusion and Outlook 564
References 564
Globally Optimal Finsler Active Contours 565
Introduction 565
Background 566
Convex Formulation of Finsler Active Contours 568
Results 571
Histology Segmentation 571
Bone Segmentation 571
Run-Time Performance for the Chan-Vese Model 573
Conclusion 573
References 574
Author Index 575