SymmNeRF: Learning to Explore Symmetry Prior for Single-View View Synthesis

AbstractWe study the problem of novel view synthesis objects from a single image. Existing methods have demonstrated potential in single-view synthesis. However, they still fail to recover fine appearance details, especially self-occluded areas. This is because only provides limited information. We observe that man-made usually exhibit symmetric appearances, which introduce additional prior knowledge. Motivated by this, we investigate performance gains explicitly embedding symmetry into scene representation. In this paper, propose SymmNeRF, neural radiance field (NeRF) based framework combines local and global conditioning under introduction priors. particular, SymmNeRF takes pixel-aligned image features corresponding as extra inputs NeRF, whose parameters are generated hypernetwork. As conditioned on image-encoded latent codes, thus scene-independent can generalize new scenes. Experiments synthetic real-world datasets show synthesizes views with more details regardless pose transformation, demonstrates good generalization when applied unseen objects. Code available at: https://github.com/xingyi-li/SymmNeRF.KeywordsNovel synthesisNeRFSymmetryHyperNetwork