A new 3D segmentation method for region-based video coding

A 3D segmentation is important for finding correspondences in region-based video coding. One approach is to extract 3D connected volume from an image sequence. However, this approach often generates undesirable 3D volume because of object´s fast movement, noise, side effects from pre-processing result and newly appeared objects. In order to avoid these undesirable results, we propose a new 3D segmentation method for extracting virtual 3D connected volume instead of real 3D connected volume. Our approach consists of intraframe mode and interframe mode segmentation. For intraframe mode, we use a modified 2D watershed algorithm on a simplified image. For interframe mode, we detect frame difference from two images. Then, previous segmented results are projected into the area where no frame difference is occurred. For the other area, we extract markers and find correspondences. To reduce search area for finding correspondences, we use three constraints like spatial, temporal and smoothness constraint