出版社：The Institute of Image Information and Television Engineers
摘要：We propose a new, simple, and effective image-based rendering technique for synthesizing images of novel views from an image cluster, a large volume of precaptured images randomly distributed in a bounded region with known camera positions and orientations. Unlike methods that are based on sampling, and constructing and resampling the plenoptic function, our method uses a select and warp approach. We use scene depth information as additional input to render novel views from an image cluster that has fewer images than that used in plenoptic function based approaches. Furthermore, unlike the view interpolation approach that typically uses a small number of input images our method deals with large number, typically from several hundred to several thousand. Our method involves finding the optimal pixel from the input image cluster on the assumption that as the image is dense enough and as each pixel needed is visible from at least one of the input images, no occlusion occurs.
