Distributed Depth Video Coding Based on Compressive Sensing and Gaussian Mixture Models

Abstract

In this paper, we present a original depth video coding method based on Gaussian mixture models (GMM) and compressive sensing (CS). depth video sequences firstly are divided into a set of group pictures (GOP) which have only one key frame and eight Wyner-Ziv frames. The conventional coding method, i.e. H.264 Intra coding is chosen to compress the key frame and the reconstructed frames are used for side information (SI). For one GOP, we acquire differences between the SI of key frame and eight Wyner-Ziv frames. Then, we utilize the CS to compress those differences in the temporal direction to obtain one compressive sensing depth differences (CSDD). The CSDD is quantized by our product vector quantizer based on GMM. At decoder, we reconstruct Wyner-Ziv frames and merge the key frame to form output videos. Experiment results show that our method is regarded as an effective coding method.