A-NeRF: Articulated Neural Radiance Fields for Learning Human Shape, Appearance, and Pose

While deep learning reshaped the classical motion capture pipeline with-forward networks, generative models are required to recover fine alignment iterative refinement. Unfortunately, the existing models are usually-crafted or learned in controlled conditions, only applicable to limited. We propose a method to learn a generative neural body model from monocular videos by extending Neural Radiance Fields (NeRFs). We them with a skeleton to apply to time-varying and articulated motion. A insight is that implicit models require the inverse of the forward used in explicit surface models. Our reparameterization defines latent variables relative to the pose of body parts and thereby ill-posed inverse operations with an overparameterization. This learning volumetric body shape and appearance from scratch while refining the articulated pose; all without ground truth labels for, pose, or 3D shape on the input videos. When used for-view-synthesis and motion capture, our neural model improves accuracy on datasets. Project website: https: //lemonatsu. github. io/anerf/.