Adaptive compressed sensing for depthmap compression using graph-based transform

In this paper we present an adaptive compressed sensing (CS) framework for depth map compression using a family of graph-based transforms (GBT). To improve overall performance, we propose a greedy algorithm that selects for each block a GBT minimizing a metric, based on average mutual coherence, that takes into consideration both the edge structure of the block and the characteristics of the CS measurement matrix. This algorithm uses a low-complexity estimate of the mutual coherence, so that explicit construction of the GBT at the encoder is not required in the iterative process. As compared to coding using H.264/AVC, the proposed approach applied to intra-frames shows an average of 39 % bitrate savings or 3.8 dB PSNR gain for views rendered using a depth image based rendering (DIBR) technique.