Prediction-Oriented Disparity Rectification Model for Geometry-Based Light Field Compression

Abstract

The geometry-based light field compression approach aims to use the inherent disparity cue provided by the light field to reduce the redundancy of data and improve immersive media broadcasting. Some methods, including our previous one, transmit explicit disparity maps along with sparsely selected views and then exploit disparity-based view prediction to remove correlations in the angular domain. However, these methods neglect two important issues: the 4D correspondence between the disparity array and view array and the practical non-uniform disparity distribution in captured light fields. As a result, the prediction potentiality of disparity maps is heavily restricted, and the limited disparity-based prediction accuracy degrades the final compression performance. In this paper, we focus on these two issues of estimated disparity maps and propose a prediction-oriented disparity rectification (PoDR) model. First, based on our previous disparity estimation, we propose to utilize the 4D structural prior of light fields to refine the estimated disparity map. The 4D correspondence between a disparity array and a view array is enhanced, leading to higher prediction accuracy and lower bit costs. Second, for the refined disparity array, we propose a variable stride-based determination algorithm to obtain the practical non-uniform disparity distribution in captured light fields. Specifically, the angular distance between each pair of disparity maps is efficiently derived by the approximate solution based on the gradient variation of the prediction errors. By combining these two modules, the proposed PoDR model improves the overall compression performance compared with our previous work. Furthermore, we verify that the proposed method can obtain better light field fundamental capability (e.g., refocusing) than state-of-the-art methods.