Reproduce “Learnable Multi-level Discrete Wavelet Transforms for 3D Gaussian Splatting Frequency Modulation” 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
Hung Nguyen, An Le, Truong Nguyen
The research introduces a framework that uses learnable multi-level Discrete Wavelet Transforms (DWT) to modulate frequency content during 3D Gaussian Splatting (3DGS) training. This approach reduces the number of Gaussian primitives by up to 100K compared to vanilla 3DGS, while maintaining competitive rendering quality through a simplified, stable coarse-to-fine optimization.
Key insights
- A deeper, multi-level frequency modulation curriculum, achieved through recursive DWT decomposition, is effective in preventing premature Gaussian densification in 3DGS.
- Simplifying learnable DWT filters to a single scaling parameter for high-pass filters significantly mitigates "gradient competition" between spatial reconstruction and frequency-domain regularization, leading to more stable and efficient optimization.
- There is a trade-off between Gaussian reduction and rendering quality when increasing DWT levels, as overly coarse initial reconstructions might introduce artifacts and slightly hinder reconstruction quality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.