MAGENTA: Magnitude and Geometry-ENhanced Training Approach for Robust Long-Tailed Sound Event Localization and Detection

2025-09-19

AI Summary

In real-world scenarios, long-tailed class distributions severely degrade the performance of deep learning-based sound event localization and detection (SELD), as standard regression losses inherently bias optimization toward high-frequency classes, undermining modeling fidelity for rare events. To address this, we propose MAGENTA—a novel method that, for the first time, unifies magnitude (radial) and direction (angular) regression errors within an interpretable vector space. MAGENTA introduces a rarity-aware geometric decomposition loss, grounded in physical principles, to explicitly guide optimization and enhance model sensitivity and robustness to infrequent events. Extensive experiments on realistic long-tailed SELD benchmarks demonstrate that MAGENTA significantly improves both localization accuracy and event detection F1-score. This work establishes the first geometric error-decoupling optimization framework tailored for long-tailed acoustic perception tasks.

Technology Category

• Machine Learning: Multi-class/Multi-label Learning & Extreme Classification
• Search and Optimization: Learning to Search
• Natural Language Processing: Learning & Optimization for NLP

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs

Abstract

Deep learning-based Sound Event Localization and Detection (SELD) systems degrade significantly on real-world, long-tailed datasets. Standard regression losses bias learning toward frequent classes, causing rare events to be systematically under-recognized. To address this challenge, we introduce MAGENTA (Magnitude And Geometry-ENhanced Training Approach), a unified loss function that counteracts this bias within a physically interpretable vector space. MAGENTA geometrically decomposes the regression error into radial and angular components, enabling targeted, rarity-aware penalties and strengthened directional modeling. Empirically, MAGENTA substantially improves SELD performance on imbalanced real-world data, providing a principled foundation for a new class of geometry-aware SELD objectives. Code is available at: https://github.com/itsjunwei/MAGENTA_ICASSP

Problem

• Addresses class imbalance in sound event detection
• Reduces bias against rare sound events
• Improves directional modeling in SELD systems

Innovation

• Geometrically decomposes regression error into components
• Enables rarity-aware penalties and directional modeling
• Unified loss function counteracts bias in vector space