The authors introduce a method called BANF (Band-limited Neural Fields) that enables frequency decomposition of neural field signals. Unlike prior approaches that rely on heuristics or specialized network architectures, BANF achieves this through a simple modification to the training process.
The key insight is that regularly sampling a neural field and then interpolating the samples using a band-limited kernel can be seen as a low-pass filtering operation. The authors leverage this to derive a cascaded training scheme that decomposes the neural field signal into multiple frequency bands.
The authors demonstrate the effectiveness of BANF across various applications, including 2D image fitting, 3D shape fitting with signed distance field supervision, and 3D shape recovery from inverse rendering. BANF outperforms prior methods in terms of reconstruction quality, especially at coarser scales, by effectively mitigating aliasing artifacts.
The authors also provide ablations to show the versatility of BANF, as it can be applied to different neural field representations without requiring modifications to the underlying architecture.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Ahan Shabano... at arxiv.org 04-22-2024
https://arxiv.org/pdf/2404.13024.pdfDeeper Inquiries