AN ADAPTIVE TRANSFER SCHEME BASED ON SPARSE REPRESENTATION FOR FIGURE-GROUND SEGMENTATION

Figure-ground segmentation benefits lots of tasks in the field of computer vision. Exemplar-based approaches are capable of performing segmenting automatically without user interaction. However, most of them adopt fixed parameters for all the target images, which blocks their segmentation performances. We present a novel sparse representation based transfer scheme to gain adaptive parameters automatically. The proposed scheme transfers the segmentation masks of some windows from training images to obtain the soft mask of the target window from any given test image, when the target window can be represented by the linear combination of those windows. On the challenging PASCAL VOC 2010 segmentation dataset, experimental results and comparisons with the state-of-the-art methods show the effectiveness of the proposed scheme.