Neural Fourier Filter Bank

Figure 1. Teaser -We propose neural Fourier filter bank to perform spatial and frequency-wise decomposition jointly, inspired by wavelets. Our method provides significantly improved reconstruction quality given the same computation and storage budget, as represented by the PSNR curve and the error image overlay. Relying only on space partitioning without frequency resolution (InstantNGP) [37] or frequency encodings without space resolution (SIREN) [47] provides suboptimal performance and convergence. Simply considering both (ModSine) [34] enhances scalability when applied to larger scenes, but not in terms of quality and convergence.