The article, that was published in the prestigious journal IEEE Transactions on Pattern Analysis and Machine Intelligence (ranked 2nd out of 85 / 206 in the field of computer science / electronics), presents a theoretical base for combining discriminative and reconstructive subspace methods. It enables robust data classification even if the information is deformed. The method is general and its usability was successfully demonstrated on different classification and regression problems.
COBISS.SI-ID: 5235540
The article presents new methods for weighted and robust learning of subspace representations of images. They enable subspace learning even in non-ideal conditions (when the images are noisy, or contain occlusions, reflections, etc.).
COBISS.SI-ID: 5898836
The article introduces an incremental method for subspace learning that enables continuous open-ended updating of the current representations while considering incoming data. In this way it abandons the traditional computer vision dichotomy of a learning stage and a recognition stage, and combines both stages into an interleaved process of recognition and incremental improvement of the current model.
COBISS.SI-ID: 6201940
In cooperation with the partners on FP5 and FP6 EU projects (Cognitive Vision Systems – CogVis and Cognitive systems for Cognitive Assistants – CoSy), we have developed a novel method for image categorization, which was published in the most cited journal in the field of computer science (ranked number 1 among 85). The method represents a big step forward in categorization algorithms development and has become one of the most used and cited methods in the field. The method is based on local patches and their constellations, which are used for modelling object categories, and successfully encompasses the diversity of exemplars of each class. This enabled a qualitative leap from the traditional recognition of (previously seen) object examplars to the recognition of (possibly very diverse) object categories.
COBISS.SI-ID: 6343252
The article gives an applicative research result, our method for determining the physical location of the optical center for an arbitrary camera.
COBISS.SI-ID: 5332820