Recognition of dexterous manipulations from time-varying images

This paper presents a method for recognizing dexterous manipulation actions that we usually perform using our hands. Our method is based on model representation using spatio-temporal vector fields and a spotting algorithm that gives segmentation-free and frame-by-frame recognition. We propose a multiview motion model that is composed of several standard sequence patterns made from different viewpoints. Results indicate that our method recognizes dexterous manipulations correctly, even if the viewing direction of the input images deviates from those of standard sequence patterns