Side Information Generation For Distributed Multi-view Video Coding

Distributed Video Coding (DVC) is a new concept based on Slepian-Wolf lossless and Wyner-Ziv lossy compression Theorems, which establish that finite-alphabet random sequences can achieve the joint entropy or achievable region bound for conventional joint encoding, with separate encoders and joint decoders. It makes possible that the side information is available at the decoder regardless of the encoder's access to side information. That is DVC could convert a bulk of computations to the decoder and enable the extensive applications.On the other hand, Multi-view techniques have been researched in the past, both for coding and for camera interpolation, since they allow creating views from virtual (non-existent) cameras, or what is called Free Viewpoint Navigation of scenes given only recordings from a few cameras. The objective of Distributed Multi-view Video Coding (DMVC) is to efficiently encode different video streams, but exploiting the possible redundancies at the decoder, thus obtaining benefits inherent to DVC like lower encoding complexity, embedded error resilience or the fact that no connection is necessary between the different cameras. In this paper, we adopt the idea of constrained relaxation for DMVC.We present a novel sub-graphs matching based framework, which devotes image segmentation and graph matching to generate inter-view correlated side information without knowing the camera parameters. Moreover, the multiple representations of images, such as point features and graph representations are incorporated to constitute more distinctive feature constraints. The sparse data as a good hypothesis space aim for best matching optimization of inter-view side information with compact syndromes, from inferred relaxed coset. The plausible filling-in from a priori feature constraints between neighboring views could reinforce a promising compensation to inter-view side information generation for joint multi-view decoding. In order to improve the stability of feature matching and accuracy for 3D viewpoint modeling, PCA-SIFT and TPS (thin plate spline) are adopted in this scheme.Four main steps involve in this scheme. The goal of segmentation at the decoder in step one is to split each image into regions that are likely to contain similar disparities that make a promising compensation for separated regions. Features are efficiently matched in step two by identifying the nearest neighbor keypoint that has the minimum Euclidean distance for the invariant descriptor vector based on PCA-SIFT. In step three, inter-view side information, from the left and right views of the WZ frame, are obtained by the geometric transform, TPS warping. Finally, a view fusion method is used to generate the inter-view side information in step four. The final side information are obtained through fusion method between temporal and inter-view side information. The experimental results show high precision for objects with high motion and improvement in the rate-distortion performance. ?...