A Deep Learning-based 6DoF Video Synthesizing Method Using Instant-NGPs

This paper introduces a new method for synthesizing six degree of freedom (6DoF) videos using neural radiance fields, which allows training from plain 2D images to render 3D scene at arbitrary viewports. Neural network model representing a previous timepoint is fine-tuned to train models for subsequent timepoints. Additionally, instant neural graphics primitives (Instant-NGP) is applied for speed improvement. The proposed method achieved both improved objective quality for same number of training iterations and enhanced consistency between frames. Furthermore, it shows superiority over other methods for generating 6DoF videos, in terms of quality and time efficiency.