Half-day:

1. Computational Symmetry: From Symmetry Group Theory to Applications and Quantitative Validations in Computer Vision and Pattern Recognition Research
2. Generalized Principal Component Analysis
3. Differential Techniques for Analysis and Synthesis in Vision and Graphics
4. Workitorial on Vision of the Unseen
5. Image Processing Techniques for Face-based Biometrics

Full Day:

1. Survey and Recent Advances in Image Registration and Fusion
2. Distributed Vision Processing in Smart Camera Networks

Half-day:

1. A Gentle Introduction to Bilateral Filtering and its Applications
2. Theory and Methods of Light-Field Photography

Full Day:

1. Computer Vision and Image Analysis in the Study of Master Drawings and Paintings

Description:

Relevance to Computer Vision:

Symmetry plays an essential role at all levels of human as well as machine perception. This tutorial is both relevant and timely for computer vision research. Even though the topic of symmetry has existed in computer vision for almost four decades, few computational tools are readily available for real world problems. A recent surge of interest in symmetry detection for both computer vision and computer graphics has underlined, once again, its importance and challenges for machine intelligence. It is time for us to take a closer look at what has been done and where we are heading from here. This tutorial will provide a succinct yet multi-faceted review for the attendees.

The goals of this tutorial are twofold:

1. To provide an accessible and concise summary of the mathematical theory for symmetry: group theory. In particular, we will cover discrete and finite symmetry groups and finitely-generated crystallographic groups.

2. To demonstrate, through multiple concrete examples, the power of using computational symmetry and symmetry groups to solve old and novel computer vision/graphics problems.

Technical content:

• Essence of Regularity and Symmetry: rising from real world problems
• Symmetry Groups: categorization and organization of symmetries -- a formal definition
• Computational Symmetry in Computer vision: Literature Review
• Computational Symmetry: Problem Formalization and Computation
• Quantified, continues versus discrete symmetry and their Applications
  • Symmetry As a Continuous Feature (revisit and a closer look)
  • Wallpaper symmetry groups and their applications in texture analysis, gait recognition and beyond
  • Regularity discovery as a higher order correspondence problem
  • Facial Asymmetry as a biometric, facial asymmetry for expression and gender classification
  • Symmetry in 3D reconstruction for computer vision
• State of the art Symmetry detection algorithms and Quantitative Evaluations of their performances
  • Why symmetry detection is interesting for computer vision/human?
  • where we started and where are we now?
  • How to evaluate the quality of symmetry detection algorithms: test images, ground truth, evaluation functions
• Summary and Conclusion:
  There is a LONG history of computational symmetry in computer vision. There is also a recent surge of interest in computational symmetry in computer vision/graphics. Where will this subfield go from here?
  
  • Theoretical directions
  • A uniform taxonomy
  • A list of un-answered questions
  • Image test sets
  • Ground truth
  • Sharing of code and data

Description: Over the past two decades, we have seen tremendous advances on the simultaneous segmentation and estimation of a collection of models from sample data points, without knowing which points correspond to which model. These advances have been motivated and constantly driven by numerous potential applications in machine learning, computer vision, image processing, systems theory, robotics, and more recently, also in biological systems. Most existing segmentation methods treat the data segmentation problem as "chicken-and-egg problem." This is because in order to estimate a mixture of models one needs to first segment the data. Conversely, in order to segment the data one needs to know the model parameters. Therefore, data segmentation is usually solved in two stages (1) data clustering and (2) model fitting, or else iteratively using, e.g. the Expectation Maximization (EM) algorithm.

This tutorial will show that for a wide variety of data segmentation problems (e.g. mixtures of subspaces, mixtures of fundamental matrices/trifocal tensors, mixtures of linear dynamical models), the "chicken-and-egg" dilemma can be tackled using an algebraic geometric technique called Generalized Principal Component Analysis (GPCA). The main idea behind GPCA is to eliminate the data segmentation step algebraically and then use all the data to recover all the models without previously segmenting the data as follows:
1. Fit a set of polynomials to all data points, without clustering the data
2. Obtain the model parameters for each group from the derivatives of these polynomials.
The tutorial will include several applications of GPCA to computer vision problems such as image/ video segmentation, 3-D motion segmentation, and dynamic texture segmentation.

Description: Differential representations are becoming popular in image/video manipulation as well as surface representation and processing. Differential operators such as image gradients contain important visual information, thereby allowing for gradient manipulations in synthesis applications to achieve globally smooth local changes in images and videos. Similarly, analysis of estimated surface gradients in photometric stereo/shape from shading is important for shape extraction. The last decade has seen a tremendous interest in gradient domain manipulation techniques for applications in vision and graphics including retinex, high dynamic range (HDR) tone mapping, image fusion (mosaics), image editing, image matting, video synthesis, texture de-emphasis and 3D mesh editing. These techniques either estimate or manipulate gradients of single image/multiple images/video/surfaces and reconstruct images/video/surfaces from the manipulated gradient fields.

Reconstruction from gradient fields, itself, has a long history in computer vision, going back to the work in Photometric Stereo, Shape from Shading and brightness constancy. In this course, we address the theoretical aspects of curl, divergence and integrability of vector fields, relevant to vision and graphics problems. We discuss scenarios where it is beneficial to operate on gradients than image intensities for image understanding, manipulation and synthesis. We review differential techniques; address issues involved in 2D and 3D reconstructions from gradients, discuss implementations/numerical methods and give in-depth technical insight into the modern applications that exploit these manipulations. The participants will learn about topics for extracting scene properties for computer vision as well as image manipulation methods for generating compelling pictures for computer graphics, with several examples. We hope to provide enough fundamentals to satisfy the technical specialist as well as tools/software’s to aid graphics and vision researchers, including graduate students.

Description: This workitutorial will combine a tutorial and workshop exploring the many facets of vision and pattern recognition to "see" what humans cannot. Topics to be

discussed include steganalysis, steganophy, invisible watermarking, forgery/manipulation detection, sensor fingerprinting, image authentication, data hiding and fusion of these topics.

Description: The tutorial will consists of two sessions devoted to the description of the basic techniques related to face recognition. The lectures will provide a comprehensive outline of face-based biometrics, its relation to biological systems (the psychophysics of the human visual system), including the existing applications and commercial systems. The lectures will then provide an in-depth analysis of the state-of-the-art algorithms for face-image analysis including: face detection and tracking, landmark localization, feature extraction, face representation and classification.

The lectures will devote most of the time to the image processing aspects of the recognition process, rather than on the classification itself. Therefore the most relevant issues and problems will be raised, providing practical solutions and algorithms responding to them. Particular attention will be devoted to the most advanced issues in the field and the current approaches presented in the literature. Finally, the tutorial will present two relevant and novel issues: the use of face image sequences for exploiting the time domain, the extension to 3D face analysis, and the categorization of facial expression to infer human emotions.

Description: This short course covers recent advances in image registration and fusion. Image registration is the process of establishing correspondence between two or more images of a scene and is required in many computer vision applications. Image fusion is the process of combining information in registered images to facilitate analysis of the images. The course covers fundamentals of image registration and fusion and reviews recent advances in the field. The materials are presented at a level understandable by attendees with some computer vision background but no prior experience with image registration and fusion; however, since most materials delivered in the course are recently published work, the course is anticipated to also be of interest to attendees who have worked in image

registration and fusion in the past. The image registration materials to be covered include: 1) robust image features that remain invariant under affine transformation, 2) information theoretic similarity measures that are suitable for rigid and nonrigid registration, 3) unbiased nonlinear models that are suitable for nonrigid image registration, and 4) registration of image structures without feature correspondence. The image fusion materials will include: 1) fusion of blurred images, 2) fusion of multi-focus images, 3) fusion of multi-exposure images, and 4) image fusion for super-resolution. This is a full-day course with presentations covering the steps in image registration and fusion, though, each presentation will be self-sufficient so attendees interested in only one topic can sit in on the course and follow the materials.

Description: Distributed vision networks is a multi-disciplinary research field that defines rich conceptual and algorithmic opportunities for the fields of computer vision, signal processing, pervasive computing, and wireless sensor networks. It also creates opportunities for design paradigm shifts in these fields given its emphasis on distributed and collaborative fusion of visual information, enabling researchers in the named areas to participate in the

creation of novel smart environment applications that are interpretive, context aware, and user-centric in nature. This course offers a perspective of the various methodologies based on the flexibilities and tradeoffs introduced by distributed vision sensing and processing. As a result of providing such perspective, the course aims to encourage participation of vision researchers in developing novel algorithms based on the potentials of distributed smart cameras.

Description: The course is split in two parts. The first part introduces the basics of bilateral filtering. We begin with the definition of the bilateral filter as a refinement upon the Gaussian blur. We then present direct applications of the filter: denoising, texture manipulation, and tone mapping. We end with the relationship between the bilateral filter and

other known image-processing techniques such as PDEs. The second half presents more advanced concepts. We describe efficient implementations of the bilateral filter. Then, we expose variants and their applications such as cross bilateral filtering and flash/no-flash photography. Finally, we discuss the limitations of bilateral filtering and ways to address them. We conclude with a Q&A session.

Description: A central area of research in computational photography is light-field capture and rendering. This includes a number of imaging techniques used to record a discrete representation of all rays in 3D space, the full 4D radiance. Compared to that, conventional photography captures only 2D images. In order to multiplex 4D radiance onto a 2D sensor, light-field photography demands more sophisticated optics and technology. In addition, rendering methods are more complex, implementing 2D projections of the 4D radiance, to form an image. This tutorial approaches the problem of light-field analysis in a rigorous mathematical way. This avoids common misunderstandings, and often

leads to solutions in unexpectedly direct way. In this sense the theoretical approach is simpler than the general "descriptive" way of teaching light-field techniques. While focusing on theoretical understanding, we also explain practical approaches, and engineering solutions to light-field problems. As part of the course, we will show a number of working light-field cameras that implement different methods for radiance capture. The cameras that we will bring and show to the audience implement the microlens approach of Lippmann and the plenoptic camera; the MERL "heterodyning" approach; our Adobe lens-prism camera; and our new camera with a mask resembling a mosquito net in front of the main lens.

Description: This full-day tutorial will apply methods from image processing, computer vision and image analysis to problems in the history and understanding of master paintings. Some of these analysis techniques are built upon methods used in forensic image analysis of photographs but are tailored to specific contingencies of painting. Questions addressed include: How do we judge the sizes and positions of objects depicted and the geometry of structures such as architecture? Was the image created using a mechanical or

optical aid? What were the sources of illumination and their color? What form of perspective was used? What is revealed by shadows and reflections depicted within a painting? Some of the analysis techniques require nothing more than a tutored and perceptive eye; others merely a ruler and pencil; yet others require advanced statistical estimation procedures and computer analysis. This course is based almost entirely on the analysis of images, not the physical or chemical analysis of pigments and media, the purview of traditional art conservators.