Depth guided selection of adaptive region of interest for Grabcut-based image segmentation

Grabcut is an efficient image segmentation technique which facilitates easy user interaction by locating a rectangular bounding box to include the foreground objects. However, when the foreground objects exhibit similar colors to that of the background, it often fail to work to accurately classify the pixels within the interior region of the bonding box. In this paper, we propose an adaptive region of interest selection algorithm for Grabcut-based image segmentation. We first obtain an initial segmentation result by performing the Grabcut on the depth image aligned to an input color image. Then we shrink and enlarge the depth segmentation mask by using the erosion and dilation operations. We regard the outside of the enlarged mask as background pixels and regard the interior of the shrunken mask as foreground pixels. The remaining pixels are classified into the foreground objects and the background by performing the Grabcut using the four-channel Gaussian mixture model of RGB colors and depth. Experimental results show that the proposed algorithm effectively suppress the false detection of objects and improves the segmentation performance compared with the existing algorithms by adaptively selecting the region of interest.