Revisiting 3D Reconstruction Kernels as Low-Pass Filters

Shengjun Zhang, Min Chen, Yibo Wei, Mingyu Dong, Yueqi Duan

TL;DR

3D reconstruction methods like 3D Gaussian Splatting often suffer from blur or aliasing because their underlying geometric primitives act as imperfect low-pass filters. This paper introduces the mathematically ideal Jinc kernel to eliminate high-frequency leakage, alongside modulated kernels that blend ideal frequency-domain filtering with fast spatial decay. The result is a significant boost in rendering quality and anti-aliasing without sacrificing computational efficiency.

What Problem Does This Paper Address?

When we capture a continuous 3D scene using discrete 2D pixels, signal processing theory tells us that this discrete sampling creates periodic replicas of the scene's spectrum in the frequency domain. To cleanly reconstruct the original continuous scene without artifacts, we need a low-pass filter that perfectly isolates the "baseband" spectrum and suppresses these high-frequency replicas (aliasing).

While explicit 3D representation methods like 3D Gaussian Splatting (3DGS) have revolutionized real-time rendering, they don't look at the problem through a classic signal processing lens. Primitives like Gaussians, Exponential functions, or Student's t-distributions act as non-ideal low-pass filters. Because they do not cut off sharply in the frequency domain, they cause high-frequency leakage, which manifests as blurred details or aliasing artifacts during novel view synthesis.

Figure 1. Frequency analysis of 3D reconstruction kernels. The table systematically summarizes key aspects for five kernel, including point clouds, Gaussians [18], exponential function [14], student’s t distribution [53], Jinc function and modulated kernels. This comparison aims to provide a qualitative reference for evaluating the spectral behavior of different kernels in 3D reconstruction, facilitating the design of optimal kernels for anti-aliasing, high-frequency information preservation, and low-frequency structural integrity.

Methods

To address these non-ideal filtering properties, the authors introduce two novel kernel designs:

  1. The Jinc Kernel: Derived from a 3D Ideal Low-Pass Filter (ILPF), the Jinc kernel is the circular/spherical counterpart to the well-known 1D Sinc filter. It features an instantaneous drop to zero magnitude exactly at the cutoff frequency, theoretically eliminating aliasing entirely. The authors derive the complete mathematical framework for "Jinc Splatting," including ray integration, coordinate transformation, and gradient backward propagation formulas for optimization.
  2. Modulated Kernels: While mathematically ideal, the Jinc kernel decays slowly in the spatial domain ($\propto r^{-1}$). When truncated to fit GPU memory tiles, it creates blocky, rectangular artifacts. To fix this, the authors propose frequency modulation. By multiplying standard, fast-decaying kernels (like Gaussians or Student's t-distributions) by a cosine term ($\cos(f_0x)$), they shift the frequencies to create a flatter, sharper, box-like passband in the frequency domain while preserving the rapid spatial decay of the base kernel.
Figure 3. Modulation of Gaussian kernels. A proper frequency shift allows the Gaussian kernel to approximate an ideal filter.

Datasets

The authors evaluated their methods on several standard benchmarks:

  • Low-Resolution NeRF-Synthetic Dataset: Because the pure Jinc kernel has high spatial support and memory demands, the authors created 64×64 and 128×128 downsampled variants of the standard NeRF-Synthetic dataset using Lanczos interpolation to evaluate pure Jinc Splatting fairly.
  • High-Resolution Datasets: To evaluate the modulated kernels, they used standard full-resolution datasets, including Mip-NeRF 360 (7 public scenes), Tanks & Temples (2 outdoor scenes), and Deep Blending (2 indoor scenes).

Key Insights and Ideas

  • Primitives as Filters: Explicit 3D primitives are not just geometry; they are spatial reconstruction kernels whose Fourier transforms dictate how well they filter out sampling noise.
  • Point Clouds Have No Filtering Power: The paper notes that point clouds act mathematically as Delta functions. Their Fourier transform is a flat line ($1$), meaning they possess zero low-pass filtering capability, explaining their poor visual fidelity in novel view synthesis.
  • The Space-Frequency Dilemma: The ideal frequency filter (Jinc) has poor spatial decay, whereas the best spatial decay kernel (Gaussian) has a soft frequency cutoff. Frequency modulation serves as a practical bridge, giving researchers the best of both worlds.

Key Results

  • Pure Jinc Superiority: In low-resolution experiments (64×64), the ideal Jinc kernel achieved a massive 5.86 dB PSNR improvement over standard 3DGS and a 0.70 dB improvement over the state-of-the-art Student Splatting (SSS) baseline.
  • Modulated Kernel Gains: On standard high-resolution datasets, Modulated Gaussians outperformed baseline 3DGS across the board, yielding up to a 0.72 dB PSNR increase on Mip-NeRF 360 along with improved SSIM and LPIPS scores.
  • Qualitative Improvements: Visually, modulated kernels successfully resolve fine structural details and maintain background color consistency where baseline methods produce blurred textures or aliasing.

Limitations and Future Work

Despite the performance gains, the authors highlight a few remaining challenges:

  • Ringing Artifacts: Because the Jinc kernel mimics an ideal filter, it inherently suffers from the Gibbs phenomenon, which can introduce subtle "ringing" or oscillatory intensity transitions rather than smooth monotonic falloffs.
  • Truncation Artifacts: If the pure Jinc kernel is cut off prematurely due to GPU memory constraints, it breaks continuity across rendering tiles, producing visible rectangular artifacts.
Figure 2. The influence of spatial decay speed. Our Jinc-based method suffers from rectangular artifacts due to premature truncation, while Gaussians benefit from its rapid decay without such artifacts. Our Modulated Gaussian Splatting method strikes an effective balance, and achieves superior rendering performance by reconciling spatial efficiency and frequency-domain fidelity.

For future work, the authors suggest extending this frequency modulation strategy to feed-forward reconstruction pipelines and 4D dynamic scenes to enhance temporal consistency.

Summary updated

Abstract

3D reconstruction is to recover 3D signals from the sampled discrete 2D pixels, with the goal to converge continuous 3D spaces. In this paper, we revisit 3D reconstruction from the perspective of signal processing, identifying the periodic spectral extension induced by discrete sampling as the fundamental challenge. Previous 3D reconstruction kernels, such as Gaussians, Exponential functions, and Student's t distributions, serve as the low pass filters to isolate the baseband spectrum. However, their unideal low-pass property results in the overlap of high-frequency components with low-frequency components in the discrete-time signal's spectrum. To this end, we introduce Jinc kernel with an instantaneous drop to zero magnitude exactly at the cutoff frequency, which is corresponding to the ideal low pass filters. As Jinc kernel suffers from low decay speed in the spatial domain, we further propose modulated kernels to strick an effective balance, and achieves superior rendering performance by reconciling spatial efficiency and frequency-domain fidelity. Experimental results have demonstrated the effectiveness of our Jinc and modulated kernels.
BibTeX
@inproceedings{zhang2026revisiting,
  author = {Shengjun Zhang and Min Chen and Yibo Wei and Mingyu Dong and Yueqi Duan},
  title = {{Revisiting 3D Reconstruction Kernels as Low-Pass Filters}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2026},
  url = {https://openreview.net/forum?id=x4uL1S3Fkg},
}
Figure