Learning to Render Novel Views from Wide-Baseline Stereo Pairs

Figure 1 . Novel view synthesis from a single wide-baseline stereo image pair. In a single forward pass, our method maps a wide-baseline stereo image pair to features that enable fast rendering of novel views, trained using only posed multi-view images of static scenes without ground-truth or proxy geometry. We outperform all prior art on novel view synthesis from sparse observations, taking a significant step towards matching the quality of overfitting on single scenes in this challenging setting.