Infotech Oulu Distinguished Seminar Series on Computer Vision and Pattern Recognition

Visual Recognition Tutorial

Lecturer: Professor Jiri Matas, Czech Technical University, Prague, Czech Republic

Date & Time:

Wednesday, August 20, 2008, 13:15 - 15:00
Thursday, August 21, 13:15 - 15:00

Room: TS127

Description

General, robust, real-time, learnable visual recognition and categorisation for large number of classes is one of the ultimate goals of computer vision. The current state of the art includes methods that fulfill some of the requirements in some restricted domains. For example, methods for real-time recognition of a very large number of locally flat, highly textured, rigid specific objects exist (e.g. Nister & Stewenius CVPR 06); in categorisation the state of the art algorithms can handle objects from tens of (carefully chosen) categories (e.g. Opelt CVPR 06). For some classes, e.g. faces, a robust and fast solution exist in fairly general settings (Viola Jones, ICCV 01). The above-mentioned approaches all have different structure, and as yet no unifying paradigm has emerged.

In the tutorial, we will first present the state of the art on selected case studies. The methods will be dissected and their assumption and components identified. In the second part, the key components and algorithms and their will be analysed and compared in depth:

• learning methods, level of supervision
• matching and search methods, search in pose space v. correspondence
  space
• object representation: appearance, features-based, edges, holistic
• robustness to occlusion, clutter,
• indexing methods

Next, we will put the current progress into perspective by reviewing "forgotten" problems and listing open problems, such as recognition of objects without surface texture (including the world of polyhedra), of wire-like object where any local patch includes background, of semi-transparent objects etc.