Probability-Based Rendering for View Synthesis

摘要: In this paper, a probability-based rendering (PBR) method is described for reconstructing an intermediate view with steady-state matching probability (SSMP) density function. Conventionally, given multiple reference images, the synthesized via depth image-based technique in which geometric information (e.g., depth) explicitly leveraged, thus leading to serious artifacts on even small errors. We address problem by formulating process as image fusion textures of all probable points are adaptively blended SSMP representing likelihood that among input images matched. The PBR hence becomes more robust against estimation errors than existing synthesis approaches. MP steady-state, SSMP, inferred each pixel random walk restart (RWR). RWR always guarantees visually consistent MP, opposed conventional optimization schemes diffusion or filtering-based approaches), accuracy heavily depends parameters used. Experimental results demonstrate superiority over approaches both qualitatively and quantitatively. Especially, effective suppressing flicker virtual video although no temporal aspect considered. Moreover, it shown map itself calculated from our RWR-based (by simply choosing most point) also comparable state-of-the-art local stereo methods.