Learning visual categories

The primary goal of present work is to develop methods for the representation of visual information that integrates appearance, structure and motion visual cues. We believe that this integration can increase current performance of visual information categorization and recognition methods.

The ability to detect and classify objects and object categories is one of the most useful functions of our visual system. We recognize almost all visual properties of
objects and scenes at a glance. We are able to learn to discriminate between object categories (e.g. people from cars) and within them (e.g. face of father from face of brother within the category of faces). At the same time, the best algorithmic methods are far from human abilities in number of categories learned, in classification speed, in the ease and flexibility of learning new categories.

Replicating humans' abilities of learning and recognition of object categories would revolutionize our everyday life. The list of possible applications that could be developed based on more efficient object category recognition technologies would contain hundreds of items, e.g. security, personalized healthcare, personal robots, design of autonomous cars and many more.

During the research we will work on methods for representation of appearance and structure of visual information from single and multiple views, on models for the analysis and integration of different visual cues, on the application of statistical learning methods for categories and on methods for categorization of objects and actions (events).