BANF: Band-limited Neural Fields for Levels of Detail Reconstruction

TL;DR

This paper introduces a simple modification to neural fields enabling effective low-pass filtering and frequency decomposition, which facilitates level-of-detail reconstruction and efficient coarser representations for downstream applications.

Contribution

The authors propose a straightforward method to produce band-limited neural fields, allowing for frequency filtering and level-of-detail processing without extensive architectural changes.

Findings

Neural fields can be effectively low-pass filtered with a simple modification.

The method enables frequency decomposition of neural signals.

Coarser representations can be computed efficiently for level-of-detail applications.

Abstract

Largely due to their implicit nature, neural fields lack a direct mechanism for filtering, as Fourier analysis from discrete signal processing is not directly applicable to these representations. Effective filtering of neural fields is critical to enable level-of-detail processing in downstream applications, and support operations that involve sampling the field on regular grids (e.g. marching cubes). Existing methods that attempt to decompose neural fields in the frequency domain either resort to heuristics or require extensive modifications to the neural field architecture. We show that via a simple modification, one can obtain neural fields that are low-pass filtered, and in turn show how this can be exploited to obtain a frequency decomposition of the entire signal. We demonstrate the validity of our technique by investigating level-of-detail reconstruction, and showing how coarser representations can be computed effectively.