Abstract

We pose unseen view synthesis as a probabilistic tensorcompletion problem. Given images of people organized by their rough viewpoint, we form a 3D appearance tensor in-dexed by images (pose examples), viewpoints, and imagepositions. After discovering the low-dimensional latent factors that approximate that tensor, we can impute its missing entries. In this way, we generate novel synthetic views ofpeople—even when they are observed from just one camera viewpoint. We show that the inferred views are both visuallyand quantitatively accurate. Furthermore, we demonstratetheir value for recognizing actions in unseen views and estimating viewpoint in novel images. While existing methodsare often forced to choose between data that is either re-alistic or multi-view, our virtual views offer both, thereby allowing greater robustness to viewpoint in novel images.