Bit allocation and encoded view selection for optimal multiview image representation

Novel coding tools have been proposed recently to encode texture and depth maps of multiview images, exploiting inter-view correlations, for depth-image-based rendering (DIBR). However, the important associated bit allocation problem for DIBR remains open: for chosen view coding and synthesis tools, how to allocate bits among texture and depth maps across encoded views, so that the fidelity of a set of V views reconstructed at the decoder is maximized, for a fixed bitrate budget? In this paper, we present an optimization strategy to select subset of texture and depth maps of the original V views for encoding at appropriate quantization levels, so that at the decoder, the combined quality of decoded views (using encoded texture maps) and synthesized views (using encoded texture and depth maps of neighboring views) is maximized. We show that using the monotonicity property, complexity of our strategy can be greatly reduced. Experiments show that using our strategy, one can achieve up to 0.83dB gain in PSNR improvement over a heuristic scheme of encoding only texture maps of all V views at constant quantization levels. Further, computation can be reduced by up to 66% over a full parameter search approach.