STaR: Self-Supervised Tracking and Reconstruction of Rigid Objects in Motion With Neural Rendering

Figure 1: An overview of our method. Given multi-view RGB videos of a dynamic scene, STaR learns a decoupled 3D representation of the static and dynamic scene components without any human annotation, which allows it to synthesize the scene from new viewpoints at new time photorealisticly, or even animate the scene with novel trajectories.