Deep region segmentation-based intra prediction for depth video coding

Depth information plays a vital role in 3D video systems. Since the depth video has large smooth areas segmented by sharp edges, preserving the sharp edges becomes a crucial task for depth video coding. Thus, depth modelling modes (DMMs) are integrated as partition prediction tools in 3D-HEVC. However, both DMM1 and DMM4 have limitations in processing diverse depth regions. To improve the performance of intra prediction for depth video coding, a novel deep region segmentation-based intra prediction (DRSIP) mode is proposed in this paper. Compared with traditional hand-crafted partition prediction methods, the proposed DRSIP mode introduces a deep region segmentation network (DRS-Net) to directly predict the segmentation result from reference texture frame. Besides, a frame-level training strategy is developed to effectively learn both local and global information for informative edge representation. Finally, the frame-level partition results are divided into block partitions to guide the reconstruction of depth blocks. Experimental results demonstrate that the proposed method achieves significant coding gains compared with the 3D-HEVC.