Generalized plenoptic sampling
Cha Zhang, Tsuhan Chen · 2001
Image-based rendering (IBR) has become a very active research area in recent years. The optimal sampling problem for IBR has not been completely solved. In this paper, we give a complete analysis on the optimal sampling problem for the lightfield. We first show that the optimal sampling efficiency can be achieved by employing the generalized periodic sampling theory with arbitrary geometry. When there is no occlusion in the scene and the scene is Lambertian, we show that the sampling density can be twice of that when we use rectangular sampling. We then propose a general framework for IBR sampling. Begin with an over-sampled dataset, we downsample the data first in the discrete domain. To render an image from the down-sampled data, two approaches are proposed, i.e., to reconstruct the over-sampled dataset first through up-sampling and then rendering, or to use a continuous interpolation filter to calculate the desired light rays directly. Eigenfilter method is employed to design filters during down-sampling and up-sampling. We show that if the proposed approach adopts the same downsampling density as the previous work, the reconstruction filter of the proposed approach is easier to design, and the reconstructed scene has a higher quality. We then extend the optimal sampling analysis to scenes with occlusions and non-Lambertian scenes. Occlusion is considered as a masking effect between layers in the scene, which gives rise to ringing artifacts in the frequency spectrum of the scene. We show that when the non-Lambertian property itself is band-limited, the Fourier transform of the scene is also expanded by a limited amount. By decreasing the sampling density appropriately, we can avoid aliasing that happens in prior work in literature, that assumes no occlusion and Lamb...