Reproduce “Compact Neural Graphics Primitives with Learned Hash Probing” 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
Towaki Takikawa, Thomas Müller, Merlin Nimier-David, Alex Evans, Sanja Fidler, Alec Jacobson, Alexander Keller
Compact NGP combines a fast spatial hash lookup with a learned probe that adjusts a small part of each feature-table address. Across eight synthetic NeRF scenes, it averages 30.66 dB at 357 kB, versus Instant NGP’s 30.93 dB at 1000 kB, retaining similar reconstruction quality with about 2.8× less storage. In one Lego-digger setup, inference is faster but training slower.
Key insights
- The unified framework of 'lookup functions' allows for novel, systematic combinations of disparate feature grid techniques, simplifying analysis and development.
- Learned probing effectively shifts storage costs from memory-intensive floating-point parameters to more efficiently compressed integer-dominated parameters.
- The combination yields a representation that is both compact and performant, achieving superior compression-quality trade-offs while maintaining fast inference times by improving cache utilization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.