On video object segmentation using fast block-matching-based self-organizing maps

This paper proposes the video object segmentation using block-matching-based self-organization maps and a kernel function. A window covering a target object in a frame is split into units with some pixels. A vector with two elements is provided a unit to quantify its color attributes, and is presented to train a map. It is also presented to classify the units. The trained map then judges the class of the unit corresponding to the presented vector. The value of the kernel function supports the mechanism of winner search from the viewpoint of a spatial feature. Experimental results show that the proposed scheme works well for the video sequence in which the viewing direction of camera moves.