Robert Loce
Xerox Research Center Webster, Xerox Innovation Group, Xerox Corp.

Mathematical Morphology in Electronic Printing

Thursday, October 11th/08:45 - 09:35

Image processing in the field of electronic printing is inherently shape-based. Objects such as text, graphics, and photo vignettes often require modification of edge location, stroke width, and corner shape, due to a need to compensate for marking process characteristics or for observer preference. This presentation will begin by describing applications of binary operations, such as conditional erosions and dilations, used to enable printing of fine black or white lines, serifs, and corners. Appearance tuning and appearance matching algorithms will be discussed. The image class will include specialized pixel types, such as high addressable pixels, half bits, and anti-aliased pixels. Loose-fitting gray-scale morphological operators for adjusting anti-aliased line art will be presented. A considerable portion of the presentation will describe "trapping," which is one of the most commonly applied morphological operations in digital image processing, although it is not typically described as morphological. Trapping is a form of conditional dilation and erosion, where the conditioning is across multiple dimensions (color planes) and gray level.

Luc Vincent
Google, Inc.

Street View: Taking Google Maps to Street Level

Friday, October 12th/08:30 - 09:20

Unveiled in May 2007, the Street View feature of Google Maps is the result of a substantial engineering effort by a team including mechanical engineers, software engineers, UI designers, computer vision scientists and scores of others. As is true with a number of other projects at Google, the initial vision for Street View was actually provided by Google co-founder Larry Page, who personally collected a street scene video from his moving car in order to bootstrap research in this area. Turning this initial vision into a product required developing major new pieces of technology, including: a robust data collection platform (ie, a van with lots of camera equipment), a fancy system for computing accurate pose from several imperfect sensors, various software components to stitch, blend, color correct and warp collected imagery, efficient systems to manage a Gargantuan flow of data, JavaScript and Flash software components to integrate Street View to Google Maps, and many others. This presentation will go over some of these components and give the audience a peek at the Street View project from behind the scene.

Luiz Velho
Instituto Nacional de Matemática Pura e Aplicada (IMPA)

Geometric and Topological Multiresolution of N-Dimensional Solids

Friday, October 12th/14:20 - 15:10

This talk will introduce a unified framework for geometric and topological multiresolution of n-dimensional solids. The framework is based on stochastic sampling the object´s support and structuring operations using alpha-complices.

Alexandre Falcão
Instituto de Computação, Universidade Estadual de Campinas (Unicamp).

The Image Foresting Transform from the Image Domain to the Feature Space

Saturday, October 13th/08:30 - 09:20

The image foresting transform (IFT) was proposed as a framework to the design of image processing operators based on connectivity. In this framework, an image is interpreted as a graph whose the nodes are usually the pixels and the arcs are defined by an adjacency relation in the image domain. A path in the graph is a sequence of adjacent pixels and each path has a value given by a path-value function. Several operators can be specified by choice of the adjacency relation and path-value function. The IFT computes an optimum path (maximum/minimum) to each pixel irrespective to its starting node (root) and in a non-increasing/non-decreasing propagation order of path values. The result is an optimum-path forest where each tree consists of the samples "more strongly connected" to its root than to any other root in some appropriated sense. Other informations can also be assigned to each pixel - the value of the optimum path, a root label, its propagation order, a graph-cut measure, the number of descendants in the forest - and the image operator is reduced to a local processing on some of these maps. The main idea was recently extended to samples of general datasets and applied to supervised pattern classification. In image analysis, these samples may be images, object contours, or pixels for example. The samples are the nodes of the graph, each sample is represented by a feature vector and the arcs are defined by an adjacency relation in the feature space (the image domain can still be used when the samples are pixels). This lecture will further exploit this recent framework by presenting a new operator for data clustering, with applications to image segmentation. The lecture will start with motivation, definitions, and a short overview of how to choose adjacency relations and path-value functions for some image processing problems. It will then present the graph model for data clustering and its results in image segmentation. This operator extends and improves some clustering methods, revealing the relation between the popular mean-shift algorithm and the watershed transforms from grayscale markers. The lecture will finish by presenting some open problems and on-going works on the IFT.