etoilekim/INR-benchmark

Signal Dataset Loader

This repository provides a small collection of synthetic and real signals—both 2D and 3D—used for compression, reconstruction.

---

Quick Start

All classes share the call signature

(dimension, length, bandlimit, seed, generate=True, super_resolution=False, sparse=False)

• dimension – 2 or 3 (ignored when not applicable)
• length – 1000
• bandlimit – fractional control variable (0.1 – 0.9 in 0.1 increments for BandlimitedSignal, SparseSphereSignal, and Sierpinski; interpretation varies per class)
• seed – ensures deterministic generation and consistent file paths (for BandlimitedSignal and SparseSphereSignal the repository ships five predefined seeds: 1234, 2024, 5678, 7890, 7618)
• generate – True = create new signal, False = load cached .npy
• super_resolution / sparse – optional toggles (see catalog below)

---

Signal Catalog

Synthetic Signals (~1 M values each)

Class	Dim	Description
BandlimitedSignal	2D / 3D	Uniform noise passed through a circular low‑pass filter; nine preset cut‑offs yield progressively higher spatial frequencies
SparseSphereSignal	2D / 3D	Random circles/spheres occupying a fixed volume fraction; sphere radius inversely proportional to bandlimit
Sierpinski	2D	Classic Sierpinski triangle rendered at depths 0 – 9, depth = int(bandlimit*10)−1
StarTarget	2D	Star‑shaped resolution target with alternating wedges; default 40 solid wedges (80 spokes total)

Real‑World Signals

Class	Notes
RealImage	Ten DIV2K images (DIV2K/00xx{,x4}.png). super_resolution=False loads the bicubic ×4 LR image; True loads the HR counterpart
Voxel_Fitting	Stanford Dragon voxel grid. sparse=True keeps only surface voxels; False loads full occupancy. super_resolution picks a higher‑res scan
CTImage	Single axial chest CT slice (chest.png), loaded as grayscale float32

---

Adding Your Own Signal

1. Subclass the same pattern and expose a self.signal NumPy array.
2. Save deterministic outputs to <ClassName>/<seed>/ so they can be re‑loaded with generate=False.
3. Keep the in‑memory footprint under ~1 M elements for apples‑to‑apples comparisons.

---

Citation & Licensing

If you use this loader in academic or industrial work, please cite:

@article{kim2025grids,
  title   = {Grids Often Outperform Implicit Neural Representations},
  author  = {Kim, Namhoon and Fridovich-Keil, Sara},
  journal = {arXiv preprint arXiv:2506.11139},
  year    = {2025}
}

Code and synthetic assets are released under the Creative Commons CC‑BY‑4.0 license. Real images remain subject to the terms of their original datasets (e.g., DIV2K).