Light field depth from multi-scale particle filtering

Jie Chen, Lap-Pui Chau, He Li

Abstract: Rich information could be extracted from the high dimensional light field (LF) data, and one of the most fundamental output is scene depth. State-of-the-art depth calculation methods produce noisy calculations especially over texture-less regions. Based on Super-pixel segmentation, we propose to incorporate multi-level disparity information into a Bayesian Particle Filtering framework. Each pixels' individual as well as regional information are involved to give Maximum A Posteriori (MAP) predictions based on our proposed statistical model. The method can produce equivalent or better scene depth interpolation results than some of the state-of-the art methods, with possible potential in image processing applications such as scene alignment and stablization.