Signals and how to store them
Signal: a function f : Ω → ℝD from image locations to D numbers (1 for grey, 3 for colour, 384 for the feature maps in this work).
| Family | Stores | Read-out | Example |
|---|---|---|---|
| Explicit | Values at grid points | Lookup | Pixels; a 16×16 or 32×32 feature grid |
| Implicit (neural field) | Weights of a function | Evaluate the network at x | NeRF, SIREN |
| Hybrid | Spatially meaningful parts | Simple rule or tiny network | Instant-NGP; sets of primitives such as Gaussians |
Two questions for any method: how is it read at a location; how are its parameters found (per-signal optimisation, or amortised: one pass of a network trained on many inputs).
Trade-offs: explicit costs memory that grows with resolution and exists only at samples (values between them come from a chosen rule); implicit is defined everywhere but needs per-signal fitting and has opaque weights; hybrids keep interpretable parts with a cheap read-out.
Key numbers: 16×16 × 384 = 98,304 stored numbers; 32×32 × 384 = 393,216. On a 256-pixel image each 16×16 cell covers 16 by 16 pixels. 64 Gaussians with 384 features each store about a quarter of the 16×16 grid's numbers and can be read anywhere.
| Method | Read-out | Parameters found by |
|---|---|---|
| Pixel grid | Lookup | Measurement |
| NeRF, SIREN | Network at x | Optimisation per signal |
| Instant-NGP | Grid lookup + tiny network | Optimisation per signal |
| This model | Simple rule over Gaussians | Placement amortised by a network |
Defence answer: “It is not just an implicit neural representation. It is read by a simple rule over explicit parts that each sit somewhere in the image, and those parts are placed by one pass of a trained network, not fitted per image.”
Source: Xie et al., Neural Fields in Visual Computing and Beyond (2022).