We developed a method for visual tracking of objects using a novel coupled-layer visual model that combines the target’s local and global appearance. Local layer is a set of local patches that geometrically constrain the changes in the target’s appearance, while global layer probabilistically models target’s global visual properties such as color, shape and apparent local motion. Local patches are constrained by global visual properties, while those are updated from stable local patches during tracking, making this coupled constraint paradigm more robust through significant appearance changes.
COBISS.SI-ID: 8801364
We developed a framework for visual-context-aware object detection from still images. The concept is based on a sparse coding of contextual features, which are based on geometry and texture. We have evaluated our method on a novel demanding image data set and compared it to a state-of-the-art method for context-aware object detection.
COBISS.SI-ID: 7684180
We developed a novel user interface concept for camera phones. Instead of typing keywords on a small and inconvenient keypad, a user just snaps a photo of her surroundings and objects in the image become hyperlinks to information. System first matches a query image to reference panoramas depicting and then augments the query image with hyperlinks to information about preannotated object in the image. Such visualization combined with touch screen allows the user intuitive access to information.
COBISS.SI-ID: 7883604
We evaluated three types of multi-class learning strategies in a hierarchical compositional framework, namely independent, joint, and sequential training. We conclude that: 1.) Joint and sequential training strategies exert sublinear growth in vocabulary size. 2.) Training time was worst for joint training, while training time even reduced with each additional class during sequential training. 3.) Different training orders of classes did perform somewhat differently. 4.) Training independently resulted in best detection rates, but the discrepancy with the other two strategies was low.
COBISS.SI-ID: 7460180
We present a case study of context aware, on-site information retrieval using computer vision based interaction on mobile phones with the goal of facilitating information access for urban commuters. The focus is on intuitive touch-less interaction using contextual recognition of existing visual objects, such as signs and information plates. Object detection and localization is based on fast matching of local image features represented as Histogrammed Intensity Patches. We show that low level image features can be learned and organized and to optimize for discrimination between similar, poorly textured object images. The system is integrated with existing software for information retrieval and experiments show that it achieves good performance in real-life situations.