Camera Field Rendering

Previous image-based rendering approaches rely on either accurate three-dimensional geometric information or dense image samples. It is, therefore, hard to render a complex real scene due to the difficulties of the following two problems: (1) how to model and represent a complex real scene using three-dimensional geometry; and (2) how to sample a real scene densely and to compress the samples. To avoid these problems, we propose two novel interpolation techniques, which can produce reasonable rendering results for sparsely sampled real scene. The first one, which is a color-matching based interpolation, searches for a possible physical point along the testing ray using color information in nearby reference images. The second technique, which is a disparity-matching based interpolation, tries to find the closest intersection between the testing ray and the disparity surfaces defined by the nearby reference images. Both approaches are designed as backward rendering techniques and they can be combined to produce robust results. In addition, a new sampling scheme called camera field is used. A camera field can be directly acquired using a two-dimensional array of cameras mounted on the corresponding support surface. Since no resampling process is required, we can fully utilize the resolution of the cameras. More importantly, new parameterization scheme allows us to find the corresponding point among reference images directly. This makes it possible to implement the proposed two interpolation approaches in the image space, rather than in the three-dimensional world space.