Reproduce “Revisiting 3D Reconstruction Kernels as Low-Pass Filters” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

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

The authors modify reconstruction kernels in scene splatting to reshape frequency response while retaining spatial decay. They derive Jinc from an ideal low-pass filter, then apply cosine modulation to Gaussian and Student’s t kernels as an approximation. On Tanks & Temples, modulated 3D Gaussian Splatting raises PSNR from 23.14 dB to 23.86 dB; scores do not measure aliasing.

Key insights

  • Framing 3D reconstruction as a signal processing problem highlights that existing kernels act as non-ideal low-pass filters, causing aliasing due to gradual frequency roll-off.
  • The Jinc kernel represents a theoretical ideal for anti-aliasing by perfectly truncating high frequencies, though its slow spatial decay poses practical implementation challenges.
  • Modulating existing fast-decaying kernels provides a balanced approach, allowing for superior frequency-domain fidelity comparable to ideal filters without sacrificing the spatial efficiency required for practical real-time rendering.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.