An Adaptive Soft Decision Method for Depth Map Prediction

Abstract

Depth map prediction is the key point in Free Viewpoint Video technique. Among current depth map estimation approaches, combined temporal and interview prediction method is the most practical one for hardware design and real time processing. It is based on block search inter-prediction on various block size to get the best estimation result with specific cost function. Although much data processing bandwidth can be saved by reusing of hardware and computation resources in disparity vector prediction, it still needs to calculate all block size cost results, and has disparity prediction errors at object boundary or continuous areas, resulting in block effect and prediction noises in depth map. This paper presents an efficient adaptive soft decision method based on chrominance image segmentation. The best prediction block size is pre-determined before block search progress. So much calculation efforts are saved. Only specific block size computation is executed to get the best disparity vector prediction, instead of selecting the best one from all block size calculation results. Experiment results show that the adaptive soft decision method can enhance depth map quality efficiently with less prediction errors and computation cost. It is suitable for hardware realization and real time processing.