Realtime-Rendering of Dynamic Scenes with Neural Radiance Fields

Kilian L. Krause, Wieland Morgenstern, Anna Hilsmann, Peter Eisert

Abstract: Digitizing the humans and scenes around us is a challenging research topic, with results finding new and more widespread implementation in modern applications. This paper addresses the challenge of real-time rendering of dynamic 3D scenes. The proposed approach utilizes Instant Neural Graphics Primitives (iNGP), training a radiance field for each individual frame. To achieve playback rates suitable for interactive VR experiences, we developed efficient data storage and loading methods. We present a modular framework for playing real-time free-viewpoint videos on a conventional Desktop PC, to showcase research advancements of dynamic radiance fields.