Relational strategies for the study of visual object recognition [Elektronische Ressource] / vorgelegt von Erol Osman

Erol Osman

Sujets

• Gesundheit

Informations

• Publié par: ludwig-maximilians-universitat_munchen
• Publié le: 01 janvier 2008
• Nombre de lectures: 34
• Langue: English
• Poids de l'ouvrage: 4 Mo

Extrait

Aus dem Institut fur Medizinische Psychologie
der Ludwig-Maximilians-Universit at Mun chen
Vorstand: Prof. Dr. Ernst P oppel
Relational Strategies for the
Study of Visual Object
Recognition
Dissertation
zum Erwerb des Doktorgrades der Humanbiologie
an der Medizinischen Fakult at der
Ludwig-Maximilians-Universit at zu Munc hen
vorgelegt von
Erol Osman
aus
Munc hen
Jahr
2008Mit Genehmigung der Medizinischen Fakultat
der Universitat Munc hen
Berichterstatter: Prof. Dr. Ingo Rentschler
Mitberichterstatter: Prof. Dr. Peter Grafe
Priv. Doz. Dr. Oliver Ehrt
Mitbetreuung durch den
promovierten Mitarbeiter: Dr. Martin Jut tner
Dekan: Prof. Dr. med. D. Reinhardt
Tag der mundlic hen Prufung: 24. Juli 2008Contents
1 Introduction 10
2 Neuroanatomy and Neurophysiology 12
2.1 Vision as Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.1.1 Precortical Visual Processing . . . . . . . . . . . . . . 12
2.1.2 Primary Visual Cortex . . . . . . . . . . . . . . . . . . 14
2.1.3 Inferior Temporal Cortex (IT) . . . . . . . . . . . . . . 15
2.1.4 Prefrontal Cortex (PFC) . . . . . . . . . . . . . . . . . 20
2.1.5 Haptic and Visual: Multimodal Processing . . . . . . . 21
2.2 Vision as Inference . . . . . . . . . . . . . . . . . . . . . . . . 22
3 Psychophysics and Model Predictions 25
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.1.1 Common Constraints . . . . . . . . . . . . . . . . . . . 25
3.1.2 Distinguishing Principles for Models . . . . . . . . . . 26
3.2 Structural Description Models . . . . . . . . . . . . . . . . . . 29
3.3 Image-Based Models . . . . . . . . . . . . . . . . . . . . . . . 33
3.4 Ideal Observer Model . . . . . . . . . . . . . . . . . . . . . . . 39
4 Syntactic Approach 41
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.1.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.1.2 Reasons for Using the CLARET-2 Algorithm. . . . . . 42
4.1.3 Machine Vision Overview . . . . . . . . . . . . . . . . 44
4.1.4 Features and Spatial Relations . . . . . . . . . . . . . . 51
4.2 Description of the CLARET-2 Algorithm . . . . . . . . . . . . 52
4.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 52
4.2.2 CLARET-2 Algorithm . . . . . . . . . . . . . . . . . . 54
4.2.3 Model Parameters . . . . . . . . . . . . . . . . . . . . 59
4.2.4 Di erences to the Original CLARET Algorithm . . . . 62
35 Learning 3D Object Representations 64
5.1 Psychophysical Experiment . . . . . . . . . . . . . . . . . . . 64
5.1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 64
5.1.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.1.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.1.4 Learning Dynamics . . . . . . . . . . . . . . . . . . . . 75
5.2 Simulation using CLARET-2 . . . . . . . . . . . . . . . . . . . 78
5.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 78
5.2.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.2.3 Simulation Experiments . . . . . . . . . . . . . . . . . 81
5.2.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
6 Generalisation 94
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
6.2 Spatial Generalisation . . . . . . . . . . . . . . . . . . . . . . 95
6.2.1 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . 95
6.2.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
6.2.3 Correlations between Learning and Generalisation . . . 98
6.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
7 Summary 107
8 Zusammenfassung 109
A Object Views for Generalisation 111
B Description of CLARET-2 116
B.1 Input Representation . . . . . . . . . . . . . . . . . . . . . . . 116
B.1.1 Intermediary Primary Attributes . . . . . . . . . . . . 116
B.1.2 Construction of Graph Representations . . . . . . . . . 118
B.2 Partitioning Attribute Spaces . . . . . . . . . . . . . . . . . . 122
B.2.1 Attribute Space . . . . . . . . . . . . . . . . . . . . . . 122
B.2.2 Partition Representation . . . . . . . . . . . . . . . . . 122
B.2.3 Conditional Attribute Space . . . . . . . . . . . . . . . 122
B.2.4 Partitioning . . . . . . . . . . . . . . . . . . . . . . . . 123
B.2.5 Relational Extension . . . . . . . . . . . . . . . . . . . 125
B.3 Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
B.3.1 Matching Algorithm . . . . . . . . . . . . . . . . . . . 125
B.3.2 Compatibility of Rules and Part Mapping . . . . . . . 126
B.3.3 Measuring Matching Quality . . . . . . . . . . . . . . . 132
4C Data and Statistical Methods 133
C.1 Trimmed Means . . . . . . . . . . . . . . . . . . . . . . . . . . 133
C.2 Distance Measure of Answering Matrices . . . . . . . . . . . . 134
2C.3 –Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
C.4 T–Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
C.5 Time-Window for Sampling Data . . . . . . . . . . . . . . . . 135
C.6 Example Classi cation Matrices . . . . . . . . . . . . . . . . . 137
C.7 Template Matching as Similarity Measure . . . . . . . . . . . 137
D OpenInventor Programs 140
D.1 Scene Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
D.2 Position les . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
Bibliography 147
Auswahl Publikationen 161
Lebenslauf 162
5List of Figures
2.1 Reduction process applied to visual stimuli . . . . . . . . . . . 17
2.2 A model of columnar organisation in TE . . . . . . . . . . . . 18
2.3 Integrative anatomy of the macaque monkey prefrontal cortex 20
3.1 List of possible indexing primitives . . . . . . . . . . . . . . . 28
3.2 Hierarchical organisation in 3D object-centred model . . . . . 30
3.3 Possible geons and their combinations . . . . . . . . . . . . . . 31
3.4 Viewing sphere . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.5 Example of possible HMAX failing . . . . . . . . . . . . . . . 36
3.6 for need for con gurational information . . . . . . . 37
4.1 Schematic of the processing stages involved in evidence-based
classi cation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.2 Example for need for part indexing . . . . . . . . . . . . . . . 50
4.3 Rule tree generated by the CRG method . . . . . . . . . . . . 51
4.4 Construction of an adjacent graph . . . . . . . . . . . . . . . . 55
4.5 Example for CLARET-2 partitioning . . . . . . . . . . . . . . 58
5.1 Objects used in learning and recognition experiments . . . . . 66
5.2 Visualisation of the 8 viewing directions . . . . . . . . . . . . 67
5.3 The set of 22 learning views . . . . . . . . . . . . . . . . . . . 68
5.4 Priming conditions allow to investigate impact of prior knowl-
edge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.5 Visual learning within a context of supervised learning . . . . 70
5.6 Supervised learning procedure . . . . . . . . . . . . . . . . . . 71
5.7 Learning times . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.8 Estimated density functions of learning times . . . . . . . . . 74
5.9 Learning dynamics . . . . . . . . . . . . . . . . . . . . . . . . 76
5.10 Answering matrices at di erent points in time . . . . . . . . . 77
5.11 Simulation experiment . . . . . . . . . . . . . . . . . . . . . . 79
5.12 Comparison of observed and predicted classi cation matrices . 85
5.13 Correlation of observed and classi cation matrices . 86
65.14 View 5 of object 2 . . . . . . . . . . . . . . . . . . . . . . . . 92
6.1 Generalisation test . . . . . . . . . . . . . . . . . . . . . . . . 94
6.2 Performance at the spatial generalisation task compared to
learning times . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
6.3 Comparison of generalisation performance for all objects . . . 97
6.4 Estimated densities of percent correct answers . . . . . . . . . 98
6.5 Performance during generalisation for known views compared
to novel views . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
6.6 Correlations for generalisation results . . . . . . . . . . . . . . 100
6.7 Counter example for possible alignment . . . . . . . . . . . . . 101
6.8 Texture like combination of object views . . . . . . . . . . . . 102
A.1 Views of object 1 tested during spatial generalisation. Only
novel views. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
A.2 Views of object 2 tested during spatial generalisation. Only
novel views. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
A.3 Views of object 3 tested during spatial generalisation. Only
novel views. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
A.4 Previously learned views of all three objects . . . . . . . . . . 115
C.1 Illustration of sampling from answering matrices . . . . . . . . 136
C.2 Classi cationprobabilitiespredictedbycross-correlation,com-
pared to human observer data. Learning views only . . . . . . 138
7List of Tables
5.1 Analysis of behavioural learning data using a t-test . . . . . . 75
5.2 Di erencesbetweengroupsatbegin, middle, andendoflearn-
ing phase . . . . . . . . . . . . . . . . . . . . . . . . . . . .