Roy Davies was Emeritus Professor of Machine Vision at Royal Holloway, University of London. He worked on many aspects of vision, from feature detection to robust, real-time implementations of practical vision tasks. His interests included automated visual inspection, surveillance, vehicle guidance, crime detection and neural networks. He has published more than 200 papers, and three books. Machine Vision: Theory, Algorithms, Practicalities (1990) has been widely used internationally for more than 25 years, and is now out in this much enhanced fifth edition. Roy held a DSc at the University of London and was awarded Distinguished Fellow of the British Machine Vision Association, and Fellow of the International Association of Pattern Recognition. Matthew Turk is a professor and department chair of the Department of Computer Science at the University of California, Santa Barbara, California. He was named a Fellow of the Institute of Electrical and Electronics Engineers (IEEE) in 2013[1] for his contributions to computer vision and perceptual interfaces. Starting on July 1st, he will be the president of the Toyota Technological Institute at Chicago[2]. In 2014, Turk was named a Fellow of the International Association for Pattern Recognition (IAPR)[3] for his contributions to computer vision and vision based interaction.

• List of contributors xiAbout the editors xiiiPreface xv1. The dramatically changing face of computer visionE.R. DAVIES1.1 Introduction – computer vision and its origins 11.2 Part A – Understanding low-level image processing operators 41.3 Part B – 2-D object location and recognition 151.4 Part C – 3-D object location and the importance of invariance 291.5 Part D – Tracking moving objects 551.6 Part E – Texture analysis 611.7 Part F – From artificial neural networks to deep learning methods 681.8 Part G – Summary 86References 872. Advanced methods for robust object detectionZHAOWEI CAI AND NUNO VASCONCELOS2.1 Introduction 932.2 Preliminaries 952.3 R-CNN 962.4 SPP-Net 972.5 Fast R-CNN 982.6 Faster R-CNN 1012.7 Cascade R-CNN 1032.8 Multiscale feature representation 1062.9 YOLO 1102.10 SSD 1122.11 RetinaNet 1132.12 Detection performances 1152.13 Conclusion 115References 1163. Learning with limited supervisionSUJOY PAUL AND AMIT K. ROY-CHOWDHURY3.1 Introduction 1193.2 Context-aware active learning 1203.3 Weakly supervised event localization 1293.4 Domain adaptation of semantic segmentation using weak labels 1373.5 Weakly-supervised reinforcement learning for dynamical tasks 1443.6 Conclusions 151References 1534. Efficient methods for deep learningHAN CAI, JI LIN, AND SONG HAN4.1 Model compression 1594.2 Efficient neural network architectures 1704.3 Conclusion 185References 1855. Deep conditional image generationGANG HUA AND DONGDONG CHEN5.1 Introduction 1915.2 Visual pattern learning: a brief review 1945.3 Classical generative models 1955.4 Deep generative models 1975.5 Deep conditional image generation 2005.6 Disentanglement for controllable synthesis 2015.7 Conclusion and discussions 216References 2166. Deep face recognition using full and partial face imagesHASSAN UGAIL6.1 Introduction 2216.2 Components of deep face recognition 2276.3 Face recognition using full face images 2316.4 Deep face recognition using partial face data 2336.5 Specific model training for full and partial faces 2376.6 Discussion and conclusions 239References 2407. Unsupervised domain adaptation using shallow and deep representationsYOGESH BALAJI, HIEN NGUYEN, AND RAMA CHELLAPPA7.1 Introduction 2437.2 Unsupervised domain adaptation using manifolds 2447.3 Unsupervised domain adaptation using dictionaries 2477.4 Unsupervised domain adaptation using deep networks 2587.5 Summary 270References 2708. Domain adaptation and continual learning in semantic segmentationUMBERTO MICHIELI, MARCO TOLDO, AND PIETRO ZANUTTIGH8.1 Introduction 2758.2 Unsupervised domain adaptation 2778.3 Continual learning 2918.4 Conclusion 298References 2999. Visual trackingMICHAEL FELSBERG9.1 Introduction 3059.2 Template-based methods 3089.3 Online-learning-based methods 3149.4 Deep learning-based methods 3239.5 The transition from tracking to segmentation 3279.6 Conclusions 331References 33210. Long-term deep object trackingEFSTRATIOS GAVVES AND DEEPAK GUPTA10.1 Introduction 33710.2 Short-term visual object tracking 34110.3 Long-term visual object tracking 34510.4 Discussion 367References 36811. Learning for action-based scene understandingCORNELIA FERMÜLLER AND MICHAEL MAYNORD11.1 Introduction 37311.2 Affordances of objects 37511.3 Functional parsing of manipulation actions 38311.4 Functional scene understanding through deep learning with language and vision 39011.5 Future directions 39711.6 Conclusions 399References 39912. Self-supervised temporal event segmentation inspired by cognitive theoriesRAMY MOUNIR, SATHYANARAYANAN AAKUR, AND SUDEEP SARKAR12.1 Introduction 40612.2 The event segmentation theory from cognitive science 40812.3 Version 1: single-pass temporal segmentation using prediction 41012.4 Version 2: segmentation using attention-based event models 42112.5 Version 3: spatio-temporal localization using prediction loss map 42812.6 Other event segmentation approaches in computer vision 44012.7 Conclusions 443References 44413. Probabilistic anomaly detection methods using learned models from time-series data for multimedia self-awaresystemsCARLO REGAZZONI, ALI KRAYANI, GIULIA SLAVIC, AND LUCIO MARCENARO13.1 Introduction 45013.2 Base concepts and state of the art 45113.3 Framework for computing anomaly in self-aware systems 45813.4 Case study results: anomaly detection on multisensory data from a self-aware vehicle 46713.5 Conclusions 476References 47714. Deep plug-and-play and deep unfolding methods for image restorationKAI ZHANG AND RADU TIMOFTE14.1 Introduction 48114.2 Half quadratic splitting (HQS) algorithm 48414.3 Deep plug-and-play image restoration 48514.4 Deep unfolding image restoration 49214.5 Experiments 49514.6 Discussion and conclusions 504References 50515. Visual adversarial attacks and defensesCHANGJAE OH, ALESSIO XOMPERO, AND ANDREA CAVALLARO15.1 Introduction 51115.2 Problem definition 51215.3 Properties of an adversarial attack 51415.4 Types of perturbations 51515.5 Attack scenarios 51515.6 Image processing 52215.7 Image classification 52315.8 Semantic segmentation and object detection 52915.9 Object tracking 52915.10 Video classification 53115.11 Defenses against adversarial attacks 53315.12 Conclusions 537References 538Index 545

Roy Davies was Emeritus Professor of Machine Vision at Royal Holloway, University of London. He worked on many aspects of vision, from feature detection to robust, real-time implementations of practical vision tasks. His interests included automated visual inspection, surveillance, vehicle guidance, crime detection and neural networks. He has published more than 200 papers, and three books. Machine Vision: Theory, Algorithms, Practicalities (1990) has been widely used internationally for more than 25 years, and is now out in this much enhanced fifth edition. Roy held a DSc at the University of London and was awarded Distinguished Fellow of the British Machine Vision Association, and Fellow of the International Association of Pattern Recognition. Matthew Turk is a professor and department chair of the Department of Computer Science at the University of California, Santa Barbara, California. He was named a Fellow of the Institute of Electrical and Electronics Engineers (IEEE) in 2013[1] for his contributions to computer vision and perceptual interfaces. Starting on July 1st, he will be the president of the Toyota Technological Institute at Chicago[2]. In 2014, Turk was named a Fellow of the International Association for Pattern Recognition (IAPR)[3] for his contributions to computer vision and vision based interaction.

• List of contributors xiAbout the editors xiiiPreface xv1. The dramatically changing face of computer visionE.R. DAVIES1.1 Introduction – computer vision and its origins 11.2 Part A – Understanding low-level image processing operators 41.3 Part B – 2-D object location and recognition 151.4 Part C – 3-D object location and the importance of invariance 291.5 Part D – Tracking moving objects 551.6 Part E – Texture analysis 611.7 Part F – From artificial neural networks to deep learning methods 681.8 Part G – Summary 86References 872. Advanced methods for robust object detectionZHAOWEI CAI AND NUNO VASCONCELOS2.1 Introduction 932.2 Preliminaries 952.3 R-CNN 962.4 SPP-Net 972.5 Fast R-CNN 982.6 Faster R-CNN 1012.7 Cascade R-CNN 1032.8 Multiscale feature representation 1062.9 YOLO 1102.10 SSD 1122.11 RetinaNet 1132.12 Detection performances 1152.13 Conclusion 115References 1163. Learning with limited supervisionSUJOY PAUL AND AMIT K. ROY-CHOWDHURY3.1 Introduction 1193.2 Context-aware active learning 1203.3 Weakly supervised event localization 1293.4 Domain adaptation of semantic segmentation using weak labels 1373.5 Weakly-supervised reinforcement learning for dynamical tasks 1443.6 Conclusions 151References 1534. Efficient methods for deep learningHAN CAI, JI LIN, AND SONG HAN4.1 Model compression 1594.2 Efficient neural network architectures 1704.3 Conclusion 185References 1855. Deep conditional image generationGANG HUA AND DONGDONG CHEN5.1 Introduction 1915.2 Visual pattern learning: a brief review 1945.3 Classical generative models 1955.4 Deep generative models 1975.5 Deep conditional image generation 2005.6 Disentanglement for controllable synthesis 2015.7 Conclusion and discussions 216References 2166. Deep face recognition using full and partial face imagesHASSAN UGAIL6.1 Introduction 2216.2 Components of deep face recognition 2276.3 Face recognition using full face images 2316.4 Deep face recognition using partial face data 2336.5 Specific model training for full and partial faces 2376.6 Discussion and conclusions 239References 2407. Unsupervised domain adaptation using shallow and deep representationsYOGESH BALAJI, HIEN NGUYEN, AND RAMA CHELLAPPA7.1 Introduction 2437.2 Unsupervised domain adaptation using manifolds 2447.3 Unsupervised domain adaptation using dictionaries 2477.4 Unsupervised domain adaptation using deep networks 2587.5 Summary 270References 2708. Domain adaptation and continual learning in semantic segmentationUMBERTO MICHIELI, MARCO TOLDO, AND PIETRO ZANUTTIGH8.1 Introduction 2758.2 Unsupervised domain adaptation 2778.3 Continual learning 2918.4 Conclusion 298References 2999. Visual trackingMICHAEL FELSBERG9.1 Introduction 3059.2 Template-based methods 3089.3 Online-learning-based methods 3149.4 Deep learning-based methods 3239.5 The transition from tracking to segmentation 3279.6 Conclusions 331References 33210. Long-term deep object trackingEFSTRATIOS GAVVES AND DEEPAK GUPTA10.1 Introduction 33710.2 Short-term visual object tracking 34110.3 Long-term visual object tracking 34510.4 Discussion 367References 36811. Learning for action-based scene understandingCORNELIA FERMÜLLER AND MICHAEL MAYNORD11.1 Introduction 37311.2 Affordances of objects 37511.3 Functional parsing of manipulation actions 38311.4 Functional scene understanding through deep learning with language and vision 39011.5 Future directions 39711.6 Conclusions 399References 39912. Self-supervised temporal event segmentation inspired by cognitive theoriesRAMY MOUNIR, SATHYANARAYANAN AAKUR, AND SUDEEP SARKAR12.1 Introduction 40612.2 The event segmentation theory from cognitive science 40812.3 Version 1: single-pass temporal segmentation using prediction 41012.4 Version 2: segmentation using attention-based event models 42112.5 Version 3: spatio-temporal localization using prediction loss map 42812.6 Other event segmentation approaches in computer vision 44012.7 Conclusions 443References 44413. Probabilistic anomaly detection methods using learned models from time-series data for multimedia self-awaresystemsCARLO REGAZZONI, ALI KRAYANI, GIULIA SLAVIC, AND LUCIO MARCENARO13.1 Introduction 45013.2 Base concepts and state of the art 45113.3 Framework for computing anomaly in self-aware systems 45813.4 Case study results: anomaly detection on multisensory data from a self-aware vehicle 46713.5 Conclusions 476References 47714. Deep plug-and-play and deep unfolding methods for image restorationKAI ZHANG AND RADU TIMOFTE14.1 Introduction 48114.2 Half quadratic splitting (HQS) algorithm 48414.3 Deep plug-and-play image restoration 48514.4 Deep unfolding image restoration 49214.5 Experiments 49514.6 Discussion and conclusions 504References 50515. Visual adversarial attacks and defensesCHANGJAE OH, ALESSIO XOMPERO, AND ANDREA CAVALLARO15.1 Introduction 51115.2 Problem definition 51215.3 Properties of an adversarial attack 51415.4 Types of perturbations 51515.5 Attack scenarios 51515.6 Image processing 52215.7 Image classification 52315.8 Semantic segmentation and object detection 52915.9 Object tracking 52915.10 Video classification 53115.11 Defenses against adversarial attacks 53315.12 Conclusions 537References 538Index 545

Roy Davies was Emeritus Professor of Machine Vision at Royal Holloway, University of London. He worked on many aspects of vision, from feature detection to robust, real-time implementations of practical vision tasks. His interests included automated visual inspection, surveillance, vehicle guidance, crime detection and neural networks. He has published more than 200 papers, and three books. Machine Vision: Theory, Algorithms, Practicalities (1990) has been widely used internationally for more than 25 years, and is now out in this much enhanced fifth edition. Roy held a DSc at the University of London and was awarded Distinguished Fellow of the British Machine Vision Association, and Fellow of the International Association of Pattern Recognition. Matthew Turk is a professor and department chair of the Department of Computer Science at the University of California, Santa Barbara, California. He was named a Fellow of the Institute of Electrical and Electronics Engineers (IEEE) in 2013[1] for his contributions to computer vision and perceptual interfaces. Starting on July 1st, he will be the president of the Toyota Technological Institute at Chicago[2]. In 2014, Turk was named a Fellow of the International Association for Pattern Recognition (IAPR)[3] for his contributions to computer vision and vision based interaction.

• List of contributors xiAbout the editors xiiiPreface xv1. The dramatically changing face of computer visionE.R. DAVIES1.1 Introduction – computer vision and its origins 11.2 Part A – Understanding low-level image processing operators 41.3 Part B – 2-D object location and recognition 151.4 Part C – 3-D object location and the importance of invariance 291.5 Part D – Tracking moving objects 551.6 Part E – Texture analysis 611.7 Part F – From artificial neural networks to deep learning methods 681.8 Part G – Summary 86References 872. Advanced methods for robust object detectionZHAOWEI CAI AND NUNO VASCONCELOS2.1 Introduction 932.2 Preliminaries 952.3 R-CNN 962.4 SPP-Net 972.5 Fast R-CNN 982.6 Faster R-CNN 1012.7 Cascade R-CNN 1032.8 Multiscale feature representation 1062.9 YOLO 1102.10 SSD 1122.11 RetinaNet 1132.12 Detection performances 1152.13 Conclusion 115References 1163. Learning with limited supervisionSUJOY PAUL AND AMIT K. ROY-CHOWDHURY3.1 Introduction 1193.2 Context-aware active learning 1203.3 Weakly supervised event localization 1293.4 Domain adaptation of semantic segmentation using weak labels 1373.5 Weakly-supervised reinforcement learning for dynamical tasks 1443.6 Conclusions 151References 1534. Efficient methods for deep learningHAN CAI, JI LIN, AND SONG HAN4.1 Model compression 1594.2 Efficient neural network architectures 1704.3 Conclusion 185References 1855. Deep conditional image generationGANG HUA AND DONGDONG CHEN5.1 Introduction 1915.2 Visual pattern learning: a brief review 1945.3 Classical generative models 1955.4 Deep generative models 1975.5 Deep conditional image generation 2005.6 Disentanglement for controllable synthesis 2015.7 Conclusion and discussions 216References 2166. Deep face recognition using full and partial face imagesHASSAN UGAIL6.1 Introduction 2216.2 Components of deep face recognition 2276.3 Face recognition using full face images 2316.4 Deep face recognition using partial face data 2336.5 Specific model training for full and partial faces 2376.6 Discussion and conclusions 239References 2407. Unsupervised domain adaptation using shallow and deep representationsYOGESH BALAJI, HIEN NGUYEN, AND RAMA CHELLAPPA7.1 Introduction 2437.2 Unsupervised domain adaptation using manifolds 2447.3 Unsupervised domain adaptation using dictionaries 2477.4 Unsupervised domain adaptation using deep networks 2587.5 Summary 270References 2708. Domain adaptation and continual learning in semantic segmentationUMBERTO MICHIELI, MARCO TOLDO, AND PIETRO ZANUTTIGH8.1 Introduction 2758.2 Unsupervised domain adaptation 2778.3 Continual learning 2918.4 Conclusion 298References 2999. Visual trackingMICHAEL FELSBERG9.1 Introduction 3059.2 Template-based methods 3089.3 Online-learning-based methods 3149.4 Deep learning-based methods 3239.5 The transition from tracking to segmentation 3279.6 Conclusions 331References 33210. Long-term deep object trackingEFSTRATIOS GAVVES AND DEEPAK GUPTA10.1 Introduction 33710.2 Short-term visual object tracking 34110.3 Long-term visual object tracking 34510.4 Discussion 367References 36811. Learning for action-based scene understandingCORNELIA FERMÜLLER AND MICHAEL MAYNORD11.1 Introduction 37311.2 Affordances of objects 37511.3 Functional parsing of manipulation actions 38311.4 Functional scene understanding through deep learning with language and vision 39011.5 Future directions 39711.6 Conclusions 399References 39912. Self-supervised temporal event segmentation inspired by cognitive theoriesRAMY MOUNIR, SATHYANARAYANAN AAKUR, AND SUDEEP SARKAR12.1 Introduction 40612.2 The event segmentation theory from cognitive science 40812.3 Version 1: single-pass temporal segmentation using prediction 41012.4 Version 2: segmentation using attention-based event models 42112.5 Version 3: spatio-temporal localization using prediction loss map 42812.6 Other event segmentation approaches in computer vision 44012.7 Conclusions 443References 44413. Probabilistic anomaly detection methods using learned models from time-series data for multimedia self-awaresystemsCARLO REGAZZONI, ALI KRAYANI, GIULIA SLAVIC, AND LUCIO MARCENARO13.1 Introduction 45013.2 Base concepts and state of the art 45113.3 Framework for computing anomaly in self-aware systems 45813.4 Case study results: anomaly detection on multisensory data from a self-aware vehicle 46713.5 Conclusions 476References 47714. Deep plug-and-play and deep unfolding methods for image restorationKAI ZHANG AND RADU TIMOFTE14.1 Introduction 48114.2 Half quadratic splitting (HQS) algorithm 48414.3 Deep plug-and-play image restoration 48514.4 Deep unfolding image restoration 49214.5 Experiments 49514.6 Discussion and conclusions 504References 50515. Visual adversarial attacks and defensesCHANGJAE OH, ALESSIO XOMPERO, AND ANDREA CAVALLARO15.1 Introduction 51115.2 Problem definition 51215.3 Properties of an adversarial attack 51415.4 Types of perturbations 51515.5 Attack scenarios 51515.6 Image processing 52215.7 Image classification 52315.8 Semantic segmentation and object detection 52915.9 Object tracking 52915.10 Video classification 53115.11 Defenses against adversarial attacks 53315.12 Conclusions 537References 538Index 545