Intra-frame depth image compression based on anisotropic partition scheme and plane approximation

3D video can be represented by color 2D video sequence accompanied by gray-scale dense depth map (depth image) sequence. In this paper, we describe a novel method for intraframe compression of the depth modality of such representation. Our method takes into account specific features of depth images, i.e. the presence of large smooth regions delineated by sharp discontinuities (edges). For such images, conventional transformbased coders produce undesirable artifacts which impede the subsequent rendering of virtual views. Instead of block transforms, the proposed method employs horizontal-vertical anisotropic partition scheme which yields a tree-structured decomposition of non-overlapping rectangular blocks adapted to the depth map content. Each block in the decomposition is approximated by plane described by the block corner pixels. The codestream consists of the coded partition scheme and the coded error of prediction of quantized corner pixels. The scheme substantially reduces the amount of artifacts around edges and yields an improvement of several dB in PSNR for typical compression ratios compared with the best transform-based coders. Other advantages of the designed coder are its simplicity and fast decompression, and the possibility to control the ratedistortion performance.