Reproduce “SoundCam: A Dataset for Finding Humans Using Room Acoustics” 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
Mason Wang, Samuel Clarke, Jui-Hsien Wang, Ruohan Gao, Jiajun Wu
SOUNDCAM is a new, large dataset comprising 5,000 real-world room impulse responses and 2,000 natural music recordings that systematically varies human positions and identities in diverse indoor environments. The dataset enables research into localizing humans with errors as low as 13.1 cm, identifying individuals with 82% accuracy, and detecting presence with 99% accuracy using multichannel acoustic data and deep learning models.
Key insights
- Multichannel RIRs, full-spectrum information (spectrograms), and higher sampling rates (48 kHz) are crucial for optimal performance in acoustic-based human sensing tasks.
- Acoustic signatures contain sufficient information to localize humans with centimeter-level accuracy and identify individuals from a group, demonstrating the potential for privacy-preserving sensing.
- Models struggle to generalize to minor room layout changes and perform significantly worse when processing raw, unknown natural audio signals compared to controlled sine sweeps or music with known sources.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.