针对声音事件检测中罕见类别识别差的问题，提出几何分解损失方法提升尾部类召回率。
MAGENTA: Magnitude and Geometry-Enhanced Training Approach for Long-Tailed Sound Event Localization and Detection
- 将回归误差分解为活动与定位两个正交分量，独立调节检测与抑制
- 在STARSS23上使总SELD误差降低20.5%，尾部类别性能显著恢复
- 无需依赖静态权重，自适应调整难样本学习，适合长尾声学场景应用
基于深度学习的声音事件定位与检测（SELD）系统在真实世界长尾声学环境中表现严重退化。标准连续回归目标对高频类别过度偏倚，导致稀有事件系统性被低估，这种优化瓶颈我们称为检测怯懦。为此，本文提出MAGENTA（Magnitude And Geometry-ENhanced Training Approach），一种与架构无关的损失框架，将回归误差几何分解为正交的径向（活动）与角向（定位）分量。不同于依赖静态频率权重的方法，MAGENTA引入内在的、由难度驱动的退火机制。通过解耦目标，独立调控活跃检测与非活跃抑制，系统可自适应提升困难尾部类别的召回率，同时调节非活跃惩罚以防止虚假稀有事件误检。在STARSS23数据集上的评估表明，MAGENTA使聚合SELD误差相对降低20.5%，有效恢复尾部类别性能，且不损害头部类别精度。代码已开源：https://github.com/itsjunwei/MAGENTA
原文摘要 · Abstract (English)
Deep learning-based Sound Event Localization and Detection (SELD) systems suffer severe performance degradation in real-world, long-tailed acoustic environments. Standard continuous regression objectives heavily bias learning toward frequent classes, causing rare events to be systematically under-recognized, an optimization bottleneck we term detection timidity. To overcome this, we propose MAGENTA (Magnitude And Geometry-ENhanced Training Approach), an architecture-agnostic loss framework that geometrically decomposes the regression error into orthogonal radial (activity) and angular (localization) components. Unlike standard methods that rely on static frequency weights, MAGENTA incorporates an intrinsic, difficulty-driven annealing mechanism. By decoupling the objective to independently modulate active detection and inactive suppression, the system can adaptively boost recall for difficult tail classes while modulating inactive penalties to prevent spurious rare-event detections. Evaluations on the STARSS23 dataset demonstrate that MAGENTA yields a 20.5% relative reduction in the aggregated SELD error, effectively recovering tail class performance without compromising head class precision. Code is available at: https://github.com/itsjunwei/MAGENTA
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。