Reproduce “Pixel-GS: Density Control with Pixel-aware Gradient for 3D Gaussian Splatting” 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

Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, Hengshuang Zhao

Pixel-GS reconstructs scenes from photos using pixel-coverage-weighted views to guide Gaussian growth from an initial SfM point cloud. It addresses sparse regions where standard 3DGS can leave blur and thin artifacts. On the nine-scene Mip-NeRF 360 evaluation, LPIPS is 0.176 versus 0.202 for retrained 3DGS*. Pixel-only weighting performs worse on Tanks & Temples; the full model reaches 0.178.

Key insights

  • The original 3DGS density control suffered from a "dilution effect" where averaging gradients across all views, including those with minimal pixel coverage, prevented large Gaussians in sparse regions from splitting.
  • Weighting gradient contributions by the number of pixels a Gaussian covers in a view accurately reflects the true need for densification in undersampled regions.
  • "Floater" artifacts are often exacerbated by disproportionately larger gradients for near-camera objects; scaling these gradients by depth effectively mitigates them.

Summary by alphaXiv.

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