NeuMan: Neural Human Radiance Field from a Single Video

Photorealistic rendering and reposing of humans is important for enabling augmented reality experiences. We propose a novel framework to reconstruct the human scene that can be rendered with poses views from just single in-the-wild video. Given video captured by moving camera, we train two NeRF models: model model. To these models, rely on existing methods estimate rough geometry scene. Those estimates allow us create warping field observation space canonical pose-independent space, where in. Our method able learn subject specific details, including cloth wrinkles accessories, 10 s clip, provide high quality renderings under poses, views, together background. Code will available at https://github.com/apple/ml-neuman .