Fridovich-Keil et al. (2023) — ARXIV-2301.10241
Key takeaways
- Represents a
d-dimensional scene withC(d,2)planes — one per pair of axes. Static isd = 3, dynamic isd = 4, and moving between them changes a number rather than a design. 1000× compression over a full 4D grid, in pure PyTorch with no custom kernels - The factorisation is chosen for what it makes expressible. Because each plane spans a named pair of axes, dimension-specific priors have somewhere to attach — temporal smoothness on the planes involving time, multi-resolution structure on the spatial ones — and static/dynamic decomposition falls out, since the planes not involving time describe what does not move
- A linear feature decoder with a learned colour basis matches a nonlinear black-box MLP decoder. Once the representation carries the structure, the decoder need not
- Competitive and often state-of-the-art fidelity across synthetic and real, static and dynamic, fixed and varying-appearance scenes
Standing in the anthology
Read — NOTE-037.
A source of SOTA-205. This reading is where the convergence was visible: it names this paper as “the third explicit-structure replacement for a large coordinate network”, after LIT-064 and LIT-108. The #121 audit filed the practice that observation implies.
Its distinctive contributions to the practice are the strong form — put the structure in the representation and the decoder can be linear — and the demonstration that no custom kernels are required.
What it keeps from NeRF (LIT-435) is the radiance field and its supervision:
it uses NeRF’s volume-rendering sum along each ray (its appendix §6.1) and
reports on NeRF’s static synthetic scenes. What it changes is where the field
lives. NeRF’s large MLP is replaced by C(d,2) multiscale feature planes whose
interpolated features are multiplied together and read out by a linear decoder
or a small MLP.
What leaves the domain is that a factorisation indexed by the axes you have assumptions about gives each assumption somewhere to live. The record’s architecture practices reason about capacity and connectivity; a representation that makes a prior expressible is doing something a more general one cannot, at equal capacity.