Seminari dei Visiting Professors

• May 30 - 3-5 P.M.
• June 1st - 3-5 P.M.
• June 4 - 5-7 P.M.
• June 6 - 3-5 P.M.

Prerequisites

This course requires knowledge of linear algebra, basic calculus, and basic probability. Previous experience with computer vision, pattern recognition, image processing and/or computer graphics will be helpful.

Course Objectives

The ubiquity of digital cameras and the internet, coupled with advances in computer vision and graphics, have been the catalysts of computational photography, an emerging field with the goal of going beyond the limitations of conventional photography by proposing new ways in which photographs can be captured, manipulated, and organized. The general goal of this seminar is to learn how, by using computational techniques and alternative camera design, computational photography produces a richer representation of our visual world. In particular, in this course, we will study ways of manipulating and combining photographs and videos to enhance the photography experience, by touching upon areas that go from mathematical models of light to advanced photography systems and techniques. Popular algorithms will be presented, which enable the development of image analysis and synthesis tools, needed to render novel images on the computer, from previously acquired samples of the real world (images or videos).

Textbook

Notes and reading material will be distributed. We will not rely on a textbook, although I recommend the free, online:

Computer Vision: Algorithms and Applications, by Richard Szeliski, Springer, 2011 

Other optional books are listed here:

• Computational Photography: Mastering New Techniques for Lenses, Lighting, and Sensors, by Ramesh Raskar, and Jack Tumblin, http://web.media.mit.edu/~raskar/photo/ (to be published soon)
• The Art and Science of Digital Compositing, Second Edition, by Ron Brinkmann, Morgan Kaufmann, 2008
• Fundamentals of Computer Graphics, Third Edition, by Peter Shirley, Michael Ashikhmin, and Steve Marschner, A K Peters Ltd, 2009 Photography, 10th Edition, by Barbara London, John Upton, Jim Stone, Pearson, 2011
• Digital Image Processing, Third Edition, by Rafael C. Gonzalez, Richard E. Woods, Prentice Hall, 2008
• Computer Vision: A Modern Approach, by David A. Forsyth and Jean Ponce, Prentice Hall, 2003

• Giovedì 26 Aprile, ore 10-13
• Venerdì 27 Aprile, ore 10-13
• Mercoledì 2 Maggio, ore 15-17

Giovedì 26 Aprile

• Lecture 1 - Image search: the problem
• Lecture 2 - Features 1: spatio-temporal-color, first things first
• Lecture 3 - Features 2: measurements & invariance, desireable properties

Venerdì 27 Aprile

• Lecture 4 - Features 3: descriptors, to grasp the image
• Lecture 5 - Detection 1: words, lots of them
• Lecture 6 - Detection 2: to context or not to context

Mercoledì 2 Maggio

• Lecture 7 - Image search: making it work: computation and samples
• Lecture 8 - Video search: performance, what can be done & what not?

• Giovedi' 12 Aprile - ore 10-12
• Venerdi' 13 Aprile - ore 11-13
• Lunedi' 16 Aprile - ore 15-17
• Mercoledi' 18 Aprile - ore 15-17

• Introduzione
• Il linguaggio delle immagini e la relazione con il linguaggio scritto
• Il significato come contenuto
  • Modelli di codifica e trasmissione del significato
  • Il contesto ed il livello semiologico del significato
• Il significato come interpretazione
  • Il contesto e l'interprete
• Il processo di creazione delle immagini e la sua influenza sul contenuto
• Gli approcci computazionali alla descrizione del significato multimedia

Venerdì 13 Aprile

• Metodi simbolici di descrizione del significato
• Concetti
  • Concetti linguistici
  • Teoria matematica dei concetti
  • Concetti visuali
• Metodi simbolici informali
  • Annotazioni
  • MPEG-7
  • Annotazione con struttura agente-azione
• Metodi simbolici informali
  • Logica del primo ordine
  • Deduzione
  • Semantica formale

Lunedì 16 Aprile

• Logica delle descrizioni
  • Sintassi
  • Semantica
• Incertezza: probabilità e logica Fuzzy
• Logica delle descrizioni fuzzy

• Ontologie
  • Cos'è una ontologia?
  • Mondo aperto o mondo chiuso?
  • Esempio di ontologia multimedia: BOEMIE
  • Ontologie fuzzy per dati multimedia

Mercoledì 18 Aprile

• Metodi non simbolici
• L'ipotesi della similarità
• Similarità preattentiva e attentiva
• Distance nello spazio delle caratteristiche
• Distanze non-metriche: "Tversky e Tversky fuzzy"
• Sistemi "prosemantici": uso di classificatori come caratteristiche relazioni tra le immagini ed il testo che le circonda
• Semantica emergente: l'interazione con l'utente
• Uso del contesto di lettura
• Qual'è oggi lo stato della semantica?

• Venerdì 23 Marzo, ore 10-12
• Lunedì 26 Marzo, ore 15-17
• Mercoledì 28 Marzo, ore 14-16
• Venerdì 30 Marzo, ore 10-12

1. Introduction to intrusion detection. Slides.
   • Taxonomy: net/host
   • Host IDS: early approaches
   • Network IDS: feature extraction
   • Signature-based IDS:
2. Anomaly detection for network security. Slides.
   • Early approaches (packet headers, PAYL, Anagram)
   • Payload based approach
   • applications (ReMIND experiments, SCADA systems)
3. Classification for network security. Slides.
   • Early approaches
   • Why it is difficult to use classification for IDS
   • behavioral classification
   • automatic signature generation
4. Attacks against learning algorithms. Slides.
   • attack taxomony
   • attacks against anagram
   • attacks against automatic signature generation

• Lecture 1 - Aula Mocci (DIEE, padiglione A)
• Lecture 2, 3, 4 - Aula Y

Abstract

Several problems in computer and network security require the analysis of a sequence of observations over time or space. Because these sequence (time-series) can be interpreted as a "signal", many computer security problems can benefit from the theory and algorithms developed by the signal processing community, and in particular, from detection theory, which focuses on determining the hypothesis that generated the signal (e.g., normal hypothesis, or attack hypothesis).



Most of the theory and results of signal processing were obtained under the assumptions of relatively benign scenarios. In signal processing we usually assume specific properties of the signal or the noise. These assumptions--while not valid at all times--may model the system accurately enough for most practical purposes.  However, when we apply signal processing tests for computer security problems, we face an intelligent opponent who will try to exploit any of our erroneous assumptions. In this course we will study how and when to apply detection theory algorithms to computer security problems.

• Motivations and Basic Methods (June 7, 9-11 A.M.)  
  • Assessment methods and Wrappers. Slides
  • Uni-variate vs Multivariate Filter Methods. Slides

• Advanced Filters and Embedded Methods (June 8, 9-11 A.M.)  
  • Information Theoretic Filters. Slides
  • Embedded Methods. Slides

• The Variance of Feature Selection (June 9, 9-11 A.M.)  
  • Stability. Slides
  • Why does Feature Selection Work? Slides

• Probabilistic Perspectives (June 10, 9-11 A.M.) 
  • Markov Blanket Algorithms. Slides
  • A Unifying View via Conditional Likelihood.

• June 7,8,10 - room X

• June 9 - room "Mocci" (DIEE - A Building)

• Tuesday May 17th 10.00 – 12 – Topics 1-2. Slides.
• Thursday May 19th 10.00 – 12 – Topics 3-4. Slides.
• Friday May 20th 10.00 – 12 – Topics 5-6. Slides.
• Tuesday May 24th 10.00 – 12 – Topics 7-8. Slides.

Topic 2: Fragile and semi-fragile digital watermarking

Topic 3. Properties and security of digital watermarking

Topic 4. Steganography and Steganalysis

Topic 5. Source device identification based on device signatures extracted from images

Topic 6. Source device identification based on enhanced sensor pattern noise

Topic 7. Unsupervised pattern classification

Topic 8. Unsupervised image clustering based on enhanced sensor pattern noise

There are circumstances where a forensic investigator has a large set of images taken by an unknown number of unknown cameras and wants to cluster those images into a number of groups, each including the images taken by the same camera, in order to narrow down the investigation. A blind image clustering method based on the enhanced sensor pattern noise will be introduced in this topic. The core of this image clustering technique is the generic classifier introduced in Topic 7.

Corso seminariale:

• Tuesday 12 April 2011 11:00-12:30; 14-15:30
  • Recalls of probability theory.
  • Main properties of the Gaussian distribution.
  • Gaussian mixture models.
• Thursday 14 April 2011 11:00-12:30; 14-16:30
  • Concepts of Bayesian classification.
  • Density estimation.
  • Parametric and non-parametric estimators.
• Monday 18 April 2011 11:00-12:30; 14-15:30
  • The hidden Markov model.
  • Action recognition.
• Wednesday 20 April 2011 11:00-12:30; 14-16:30
  • Tracking in videos.
  • Recursive Bayesian estimation.
  • The Kalman filter.
  • Particle filters.

Objectives

This course will introduce the students to techniques for action recognition and tracking in videos. The first few lectures offer recalls of foundational topics such as probability, Bayesian classification and density estimation which will facilitate the comprehension of the following topics. Action classification covers the hidden Markov model, bag-of features approaches and conditional random fields. Tracking covers recursive Bayesian estimation and the Kalman and particle filters. The course aims to be of benefit for research students in the areas of computer vision, pattern recognition, multimedia, image and signal processing, human-computer interaction, and also computer science/computer engineering in general.

Learning modalities

Materials

A copy of the slides can be downloaded from here. The copyright of all materials stays with the copyright holders.



Topics at a glance

• Recalls of probability theory
• Concepts of Bayesian classification
• Density estimation
• The hidden Markov model 
• Action recognition
• Tracking in videos.
• Recursive Bayesian estimation.
• The Kalman filter.
• Particle filters

• Lunedì 20 September, 9-13
• Mercoledì 22 September 23, 9-13

Social media creation and use has skyrocketed in recent years and has become an indelible part of our lives -- from the way we entertain and inform ourselves to the way we communicate, socialize, or learn. More than 24 hours of new video content are uploaded on YouTube each minute, and over 100M new photos are uploaded on Facebook every day.  With the tremendous growth of online multimedia content come great opportunities but also great expectations and challenges.  Users expect images and video to be searchable as easily as text but technology has unfortunately not kept pace.

In this short course, I will present the current state of art in multimedia search, and will review the current approaches for visual content-based as well as semantic indexing and search. Emphasis will be given on a new promising direction of research, semantic concept-based retrieval, which aims to boost both the effectiveness and usability of multimedia search. I will describe techniques that leverage the computer's ability to effectively analyze visual features of images and video, and apply statistical machine learning techniques to classify and label visual scenes, objects, people, and activities, automatically.  I will also describe methods that leverage such automatically generated labels to improve the quality of multimedia indexing and search as well as to enable new applications and content monetization models.  The above approaches will be presented and demonstrated in the context of a state-of-art multimedia analysis and retrieval system developed at IBM Research.

Here is a preliminary synopsis and list of topics to be covered (still subject to change but not by much):

1. Introduction and motivation

1.1 Opportunities of multimedia analysis and retrieval

1.2 Challenges and basic problems

1.3 Example applications and demos

2. Content-based retrieval

2.1 Global Visual Features

- Color spaces and color features

        - Color feature representations (color histograms, correlograms, moments)

        - Texture features (structural, statistical, spectral)

        - Edge and shape features

2.2 Local visual features

        - Interest point detection

        - Local descriptor representation

        - Local point matching and spatial registration

2.2 Similarity measures and evaluation metrics

2.3 Video segmentation and matching

2.4 Advanced techniques of content-based retrieval

2.4 Video fingerprinting and near-duplicate detection

3. Semantic concept-based retrieval

3.1 Definitions and motivation

3.2 Semantic concept vocabulary design

3.3 Semantic concept modeling and extraction

3.4 Multi-modal fusion and semantic context exploitation

3.5 Retrieval by semantic concepts

3.6 Concept-based query expansion

4. Multi-modal video retrieval -- a case study

4.1 Speech-based retrieval

4.2 Visual content-based retrieval

4.3 Semantic concept-based retrieval

4.4 Query-dependent multi-modal fusion

4.5 Performance evaluation (TRECVID)

4.6 Demos and other applications

• Mercoledì 20 Gennaio, 11-13
• Giovedì 21 Gennaio, 11-13
• Lunedì 25 Gennaio, 11-13
• Martedì 26 Gennaio, 11-13

Over the last ten years, face processing (detection and recognition) has become an important research area in computer vision, due to its wide range of possible applications such as human-computer interaction, video surveillance systems, biometrics and content-based multimedia indexing  and retrieval.

The goal of face detection is to determine whether or not there are any faces in the image and, if present, their location. It is the crucial first step of any application that involves face processing systems. Thus, accurate and fast human face detection is the key to a successful operation.

Face detection is a challenging problem because faces highly vary in size, shape, color, texture and location. Their overall appearance can also be influenced by lighting conditions, facial expression, occlusion or facial features, such as beards, mustaches and glasses. Another challenging problem comes from the orientation (upright, rotated) and the pose (frontal to profile) of the face.

The first part of these lectures will introduce the basics of face detection and will review some of the numerous techniques proposed to counter these issues.

Face recognition has existed for more than 30 years and has been particularly active since the early 1990s. Researchers from many different fields (from psychology, pattern recognition, neuroscience, computer graphic and computer vision) have attempted to create and understand face recognition systems.

This has led to many different techniques often divided into two groups:

The second part of these lectures will address in details the most famous and successful face recognition techniques including the recent development that make use of Local Binary Patterns.

Topics:

• Venerdì 16, 14-16
• Lunedì 19, 10-12
• Mercoledì 21, 10-12
• Venerdì 23, 10-12

• Venerdì 16: Aula Mocci, Padiglione A
• Lunedì19, Mercoledì 21, Venerdì 29: Sala Riunioni, Padiglione B

This course aim to arm the participants by hand-crafing a Bayesian classifier using both generative and discriminative approaches. Advanced topics such as dealing with missing values and Bayesian network are also covered. These techniques enable one to solve many machine-learning problems where training data is an integral part of the system design. Practical applications are illustrated via multimodal information fusion in biometrics.

The course (8 lecture hours) will be roughly structured as shown below:

• Lecture 1: Decision theory, generative and discriminative classifiers, Naive Bayes. Slides.
• Lecture 2: Error prediction in multimodal information fusion. Slides.
• Lecture 3: User-specific score normalization and fusion. Slides
• Lecture 4: Quality-based fusion. Slides

Lectures 2-4 have example applications applied to biometrics.

Bibliografia

• Notes

Assignement

• Instructions.txt
• Data.txt

• Venerdì 11, 10-13
• Mercoledì 16, 11-13
• Venerdì 18, 10-13

Classifier ensembles are a growing area of pattern recognition and machine learning. This course will set up the pattern recognition background and will introduce the basics of classifier ensembles. Ensemble approaches and methods will be presented including the "classics" such as bagging, boosting, random subspace, random forest end ECOC ensembles, as well as more recent additions to the ensemble collection, e.g., rotation forest and random oracle ensembles. We will take a look at the ensemble design strategies and combination rules, and will discuss intuition and theory about why classifier ensembles work. Diversity among the classifiers in the ensemble will be explained, along with its relationship with the ensemble accuracy. Cluster ensembles will be briefly presented, and we will touch upon classifier ensembles for changing environments.

The course (8 lecture hours) will be roughly structured as shown below:

1. Pattern Recognition: basic concepts, evaluation and comparison of classifiers. Slides

2. Base Classifier Models. Slides

3. Combination Methods. Slides

4. Diversity in Classifier Ensembles. Slides

5. Building Classifier Ensembles: error correcting output codes (ECOC); classifier fusion and classifier selection; dynamic classifier selection; bagging, AdaBoost, random forests and random oracle ensembles. Slides

6. Feature Selection:  Selection and extraction of features for ensembles and from ensembles. Rotation Forest. Slides

7. Cluster ensembles: construction methods, aggregation of the individual decisions, diversity. Slides

8. Classifier Ensembles for Changing Environments: concept drift and types of changes; detecting a change; individual online classifiers; ensemble approaches for concept drift. Slides

I dottorandi che sono interessati a sostenere la prova finale possono trovarne il testo qui. La soluzione deve essere inviata via e-mail alla Prof. Kuncheva entro il 5 Ottobre. Potete trovare qui il suo indirizzo email.

• Tuesday 18, 14-16
• Wednesday 19, 14-17
• Thursday 20, 10-13

• Tue 18, Wed 19 - Aula Mocci, Padiglione A
• Thu 20 - Aula X Ingegneria (via is Maglias)

1. Data Mining for Intrusion Detection

   In this lecture, the basics in applying data mining techniques to intrusion detection will be discussed. Basic elements in building an intrusion detection system will be first provided. Then, a data mining framework for constructing features and models for intrusion detection systems will be described. Moreover, how to learn models that are cost-sensitive and how to lean attack scenarios from intrusion alert data will be discussed.

2. Accuracy and Robustness Issues in Intrusion Detection

   The first part of this lecture concerns how to learn a cascade anomaly detection system that has high detection rate and low false positive rate. Also, both a "polymorphic blending attack" that can defeat simple anomaly detection system and machine-learning based method to improve the robustness of detection models will be described. Moreover, a "noise injection attack" that can defeat syntactic learners of worm signatures will be described.

3. Botnet detection in enterprise network

   A brief background on botnets will be firstly provided. Then, three botnet detection systems (BotHunter, BotSniffer, and BotMiner) will be described. BotHunter uses "vertical correlation" to analyze events involving a host that match the life-cycle of botnet. BotSniffer uses "horizontal correlation" to detect a host or group of hosts that have synchronized communication or malicious activity traffic that suggest bot-like behavior. BotMiner is more general: it clusters hosts with similar communication traffic, clusters hosts with similar malicious activity traffic, and find the intersection of the hosts to detect bots.

• Tuesday 15, 15-18
• Wednesday 16, 10-13
• Thursday 17, 10-12

1. Document Image Understanding for Digital Libraries:

   Web-based digital libraries: rapid expansion worldwide. Automatic document image understanding state of the art

2. Document Content Extraction:

   Highly versatile document processing. Pixel-accurate segmentation and iterated classification

3. Combination of Classifers:

   There is no dominant classifier. Combinations can be better; stochastic discrimination

4. Contextual Analysis as an Aid to Recognition:

   Geometric, linguistic, and semantic context. The promise of computable semantics: reading Chess

5. Adaptive Recognition Systems I:

   Self-retraining classifiers. Autonomously adaptive classifiers

6. Adaptive Recognition Systems II:

   Exploiting isogeny (internal consistency) of documents. Towards whole-book recognition: anytime classifiers

7. Human Interactive Proofs I:

   Survey of CAPTCHAs and other HIPs. The arms race: segmentation, occlusion, and linguistics

8. Human Interactive Proofs II:

   Research frontier: exploiting style consistency and linguistic familiarity. Future directions: implicit CAPTCHAs, non-textual challenges

slides

• Wednesday 25, 15-18
• Thursday 26, 10-13
• Friday 27, 10-12

• Wed 25, Thu 26 - Aula Y
• Fri 27 - Aula X Ingegneria (via is Maglias)

1. Introduction

   Pattern recognition problem, PR system, statistical decision rules, multiple classifier paradigm, multiple classifier system architectures, score normalisation

2. Multimodal expert fusion

   Problem formulation, simple fusion rules and their relationship, error sensitivity, trained rules

3. Coping with estimation errors

   Bayesian error probability, error distribution, actual error probability, intramodal fusion architectures, diversification and marginalisation

4. Margin manipulation approaches

   Error correcting output coding approaches, serial multiple classifier systems

5. Classifier selection
6. Quality based fusion of multiple experts

slides

• Monday 24, 17-19
• Tuesday 25, 11-13
• Wednesday 26, 11-13

1. Genome evolution
2. Learning with structured outputs
3. Patterns in media content
4. The analysis of patterns

slides

• Tuesday 26, 10-12
• Wednesday 27, 11-12
• Thursday 28, 10-12
• Friday 29, 10-11

• Tue 26 - Sala riunioni (pad B)
• Wed 27, Thu 28 - Aula Y
• Fri 29 - Aula X

1. Overview of Bayesian and MDL Model Selection
2. Model and feature selection in mixture-based clustering
3. Semi-supervised Classification
4. Semi-supervised Clustering: Application to Image Segmentation
5. Advanced Applications of the EM Algorithm: Network Inference

• Wednesday 6, 10-12
• Thursday 7, 16-18
• Friday 8, 10-12

• Wed 6, Fri 8 - Aula Y
• Thu 7 - Aula X

1. Selective Ensemble
2. Learning Comprehensible Models
3. Multi-Instance Learning
4. Multi-Instance Multi-Label Learning
5. Semi-Supervised Learning
6. Exploiting Unlabeled Data in Content-Based Image Retrieval

slides

• Tuesday 15, 09-11
• Wednesday 16, 17-19
• Thursday 17, 17-19

• Tue 15 - Aula Y
• Wed 16, Thu 17 - Aula I

1. Biometric Systems Overview
2. Statistical Metrics of System Performance
3. Vulnerability of Biometric Systems
4. Liveness Assessment
5. Inference for Single Population Metrics
6. Inference for Multiple Population Metrics