Viewpoint determination of image by interpolation over sparse samples

摘要

We address the problem of determining the viewpoint of an image without referencing to or estimating explicitly the 3-D structure pictured in the image. Used for reference are instead a number of sample snapshots of the scene, each supplied with the associated viewpoint. By viewing image and its associated viewpoint as the input and output of a function, and the reference snapshot–viewpoint pairs as input–output samples of that function, we have a natural formulation of the problem as an interpolation one. The interpolation formulation allows imaging details like camera intrinsic parameters to be unknown, and the specification of the desired viewpoint to be not necessarily in metric terms. We describe an interpolation-based mechanism that determines the viewpoint of any given input image, which has the property that it fits all the given input–output reference samples exactly. Experimental results on benchmarking image datasets show that the mechanism is effective in reaching quality viewpoint solution even with only a few reference snapshots.