Reproduce “MERF: Memory-Efficient Radiance Fields for Real-time View Synthesis in Unbounded Scenes” 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

Christian Reiser, Richard Szeliski, Dor Verbin, Pratul P. Srinivasan, Ben Mildenhall, Andreas Geiger, Jonathan T. Barron, Peter Hedman

MERF (Memory-Efficient Radiance Fields) enables real-time, high-quality rendering of unbounded 3D scenes directly in web browsers. It achieves over 17x memory reduction compared to prior real-time methods while maintaining visual quality and enabling interactive frame rates on consumer hardware.

Key insights

  • A hybrid 3D voxel grid combined with 2D feature planes offers an efficient way to represent NeRFs, balancing memory use and detail preservation.
  • A novel piecewise-projective contraction function for unbounded scenes allows rays to remain straight in contracted space, simplifying ray-AABB intersection for real-time rendering.
  • Integrating quantization directly into the training process and leveraging proposal MLPs during baking helps convert the representation into a compact, high-quality format suitable for real-time browser-based rendering.

Summary by alphaXiv.

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