Reproduce “Position tracking of a varying number of sound sources with sliding permutation invariant training” 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
David Diaz-Guerra, Archontis Politis, Tuomas Virtanen
A novel training strategy called Sliding Permutation Invariant Training (sPIT) is presented to enhance the consistency of deep learning-based sound source tracking. This method significantly minimizes identity switches in dynamically varying multi-source environments, demonstrating improved tracking robustness compared to frame-level approaches.
Key insights
- Applying permutation matching over a causal temporal window effectively reduces identity switches, providing a crucial balance between frame-wise accuracy and long-term tracking consistency.
- The use of ACCDOA combined with sPIT inherently optimizes both source localization and the detection of source appearance and disappearance (misses and false positives) within a single loss function.
- sPIT maintains a computational cost comparable to fPIT, offering a robust and practical training strategy without requiring complex auxiliary networks or suffering from convergence issues seen in other methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.