Compression of PlenOctree Model Attributes Enabling Fast Communication and Rendering of Neural Radiance Fields

Abstract

Neural Radiance Fields (NeRF) have led the advancement of techniques for 3D scene representations for synthesizing views through the use of a multilayer perceptron. Its inability to achieve real-time performance for photorealistic image rendering, however, has enabled the advent of methods that speed-up rendering time at the cost of rendering quality or model complexity. In this work, we focus on the existing Plenoctree method, which possesses high rendering speed but unfortunately needs a large space for storing and transmitting its model. We address this weakness by proposing improvements to the different blocks of its pipeline and adding an efficient compression stage, without modifying the underlying representation, while maintaining high rendering quality and speed. Results over a set of test camera poses - using our methods which were obtained with data from the training and validation datasets - show that we can reduce about eight times the bit rates of the encoded models and still obtain a higher quality of the synthesized images when comparing them to the original PlenOctree models or, alternatively, a reduction of about 50 times while presenting minimal degradation for novel view synthesis.