Local Background Sprite Generation

Background modeling of video sequences can be used in many different applications. For video object segmentation it is often applied in a background subtraction method. When conventional sprites like single or multiple sprites are used a background sequence has to be reconstructed from the model. The double mapping into the coordinate system of the sprite and back can lead to severe distortion of the background model and therefore to erroneous segmentation masks. We present a novel background modeling approach that lessens distortion. These so-called local background sprites are built for every reference frame independently and fit its original size. Experimental results show that this new approach clearly outperforms conventional background sprites in terms of PSNR.