Fast View Interpolation from Stereo: Simpler can be Better

Résumé

In this paper, we propose to rely only on images to generate novel views, and recall why modeling of the complete scene is often too expensive in the context of view interpolation. We investigate ways to achieve view interpolation by mean of forward and backward mapping. We present situations in which each one requires less computations and gives better results. Contrary to what we might expect, very simple stereo algorithms can produce very convincing interpolation despite providing really bad disparity maps. We propose to explain this with a probabilistic model of depth discontinuities. We test this model on synthetic data created to fit real image statistics and compare with images widely used in stereo. In practice, forward and backward mapping methods can rely on simple stereo algorithms running in real time, to produce very good results. A sequence of real images was acquired to allow accurate comparison of interpolated images, and standard metrics are used to assess the quality.

Documents

PDF

---