Gridless 3D Recovery of Image Sources From Room Impulse Responses

Tom Sprunck, Antoine Deleforge, Yannick Privat, Cédric Foy

TL;DR

This paper introduces a "gridless" method to reconstruct the 3D positions and amplitudes of hundreds of acoustic image sources from a single room impulse response (RIR). By treating the problem as a convex optimization over the space of Radon measures—rather than a discrete grid—the authors achieve unprecedented localization accuracy (sub-millimeter and sub-degree) that bypasses the resolution limits of traditional "on-the-grid" methods.

What Problem Does This Paper Address?

When sound bounces off walls in a room, we can model those reflections as "image sources"—virtual speakers located outside the room. Mapping these sources is key to "hearing" the shape of a room, improving hearing aids, or creating realistic spatial audio for AR/VR.

Traditionally, researchers solve this by defining a discrete 3D grid and searching for sources at those specific points. However, this "on-the-grid" approach has three major flaws:

  1. The Curse of Dimensionality: To get high precision in 3D, the grid size explodes (e.g., needing over 100 million points for 1cm precision).
  2. Basis Mismatch: If a real source sits between two grid points, the algorithm struggles to represent it accurately.
  3. Peak Overlap: When reflections arrive close together in time, standard peak-picking fails.

Methods

The authors borrow from recent "super-resolution" mathematical frameworks to solve the problem in continuous 3D space.

  • Convex Relaxation: They reformulate the non-convex problem of finding discrete spikes into a convex linear inverse problem over the space of Radon measures. This is known as the Beurling-LASSO (BLASSO).
  • The Linear Operator: They derive a new operator ($\Gamma$) based on the fundamental solution of the wave equation. This operator maps a distribution of sources to the signals measured by a microphone array.
  • Sliding Frank-Wolfe (SFW) Algorithm: To solve the optimization, they use an adapted SFW algorithm. Unlike standard greedy algorithms that just add points, SFW includes a "sliding" step: after adding a new source, it jointly refines the positions and amplitudes of all currently estimated sources using a local non-convex search (BFGS). This allows the sources to "drift" to their exact continuous coordinates.

Datasets

The study relies on simulated data to establish a baseline for what is mathematically possible:

  • 200 Simulated Rooms: Randomly sized "shoe-box" rooms (lengths/widths 2–10m, heights 2–5m) with varying wall absorption coefficients.
  • Pyroomacoustics: Used to generate ground-truth image source positions and RIRs.
  • Microphone Array: A simulated 32-channel spherical array (modeled after the Eigenmike) with varying diameters.

Key Insights and Ideas

  • Gridless is Greater: By moving away from grids, the authors achieve a level of precision that would require a grid four orders of magnitude larger than what they used for initialization.
  • The Power of the "Sliding" Step: The ability to continuously refine positions during the optimization process is what allows the recovery of hundreds of sources. Previous sparse methods typically capped out at a few dozen.
  • Echo Density vs. Recovery: The method performs significantly better in larger rooms. In small rooms, "echo density" (the number of reflections arriving per millisecond) becomes so high that sources overlap too much to be separated, even with super-resolution techniques.

Key Results

The algorithm was tested on the first 50ms of RIRs (where early reflections are most distinct):

  • High Recall: In large rooms, the method recovered 94.3% of audible image sources.
  • Extreme Precision: For recovered sources, the mean angular error was just 0.38° and the mean radial error was less than 0.1 mm.
  • Robustness: The system maintains high performance at 40 dB PSNR. While performance drops at 30 dB, the authors found that simply relaxing the "success" threshold (e.g., from 2° to 6° angular error) restored the recall rate to near-noiseless levels.
  • Hardware Impact: Increasing the sampling rate (to 32 kHz) or the array diameter (to 42 cm) significantly improved the recovery of sources in dense acoustic environments.

Limitations and Future Work

While the results are a major step forward for acoustic imaging, several hurdles remain for real-world deployment:

  • Idealized Physics: The current model assumes an ideal low-pass filter for microphones and frequency-independent wall reflections. Real-world materials and hardware have complex, frequency-dependent responses.
  • Room Geometry: The study focused on cuboid (shoe-box) rooms. Non-convex or complex polyhedral rooms would introduce occlusions (sources audible to some mics but not others), which the current operator does not account for.
  • Computational Cost: While more efficient than a massive 3D grid, the iterative sliding optimization is still computationally intensive compared to simple peak-picking.

The authors suggest that future work will involve extending the framework to the Fourier and spherical-harmonic domains to handle real-world frequency dependencies.

Summary updated

Abstract

Given a sound field generated by a sparse distribution of impulse image sources, can the continuous 3D positions and amplitudes of these sources be recovered from discrete, band-limited measurements of the field at a finite set of locations, e.g., a multichannel room impulse response? Borrowing from recent advances in super-resolution imaging, it is shown that this non-linear, non-convex inverse problem can be efficiently relaxed into a convex linear inverse problem over the space of Radon measures in $\mathbb {R}^{3}$. The new linear operator introduced here stems from the fundamental solution of the wave equation combined with the receivers' responses. An adaptation of the Sliding Frank-Wolfe algorithm is proposed to numerically solve the problem off-the-grid, i.e., in continuous 3D space. Idealized simulated experiments show that the approach can recover hundreds of image sources at a rate and accuracy that are not achievable by previous methods, using a compact microphone array and source placed at random in random-sized shoe-box rooms. The impact of noise, sampling rate and array diameter on these results is also examined.
BibTeX
@article{sprunck2022gridless,
  author = {Tom Sprunck and Antoine Deleforge and Yannick Privat and Cédric Foy},
  title = {{Gridless 3D Recovery of Image Sources From Room Impulse Responses}},
  journal = {IEEE Signal Processing Letters},
  year = {2022},
  url = {https://openreview.net/forum?id=RqSX5d7SI5},
}

Loading related content...

Figure