ACCDOA: Activity-Coupled Cartesian Direction of Arrival Representation for Sound Event Localization and Detection

1. INTRODUCTION

• SELDnet uses SED and DOA targets with a combined loss, while separate task-specific networks avoid that multi-objective problem but increase complexity and network size.Keeping networks small is important for edge devices.Source: §1 · p. 1, Figure 2 · p. 2
• On DCASE 2020 Task 3, ACCDOA outperformed two-branch representations in SELD metrics with fewer parameters and state-of-the-art systems in localization error and location-dependent F-score.Source: §1 · p. 1
• ACCDOA assigns event activity to the length of its Cartesian DOA vector, letting one network solve SELD with a single target.After feature extraction, the network outputs frame-wise ACCDOA vectors for target events.Source: §1 · p. 1, Figure 1 · p. 1
• The model minimizes distance between estimated and target ACCDOA coordinates; when no event is present, loss is calculated only for activity, not DOA.Source: §1 · p. 1

2.1. Overview of NN-based SELD methods

• Physics-based DOA estimation has been integrated with neural SED, while neural-only methods perform well when labeled data are available.Examples include CRNN-based SED combined with a histogram DOA algorithm [9, 10] and neural denoising or source separation to refine intensity-vector DOA.Source: §2.1 · p. 2
• Neural SELD methods use either joint SED and DOA targets or task-specific subnetworks; cited systems include competition-leading examples of both designs.Wang et al. used a SELDnet-like architecture, while Kapka and Lewandowski decomposed SELD into four tasks.Source: §2.1 · p. 2

2.2. SELDnet

• SELDnet was presented as the first method to localize and recognize more than two overlapping events while tracking activity per frame; a CRNN embedding feeds parallel SED and DOA branches.Source: §2.2 · p. 2
• SELDnet uses BCE for SED and MSE for DOA, combining the losses linearly and searching weights to balance them during optimization.Source: §2.2 · p. 2

2.3. Two-stage method

• The two-stage method trains the SED network first, transfers some parameters to the DOA network, then trains it with masked MSE that omits inactive events.Source: §2.3 · p. 2, Figure 3 · p. 2

3.1. ACCDOA representation

• ACCDOA encodes each event class with a vector whose length represents activity and whose direction represents DOA when the class is active.Reference activity is 1 when active and 0 when inactive; active Cartesian DOAs have unit length.Source: §3.1 · p. 2, Figure 4 · p. 3
• Activity and Cartesian DOA are recovered from ACCDOA, with DOA considered only for active classes and thresholding used to determine activity.The figure illustrates activity through vector length and active-event DOA through vector direction.Source: §3.1 · pp. 2–3, Figure 4 · p. 3

3.2. SELD with ACCDOA representation

• A CRNN embedding network maps acoustic features to an embedding, which one fully connected layer uses to estimate ACCDOA vectors.The resulting SELD outputs require neither two task-specific branches nor two task-specific networks.Source: §3.2 · p. 3
• A single MSE loss trains estimated ACCDOA vectors toward their references, avoiding the multi-objective problem of balancing separate task losses.When ground truth indicates no event, the loss is calculated for activity but not DOA; activity is not interpreted as a posterior probability because BCE is not used.Source: §3.2 · p. 3, Figure 5 · p. 3

4.1. Task setups

• The 600-recording dataset split into 400 training, 100 validation, and 100 testing recordings; evaluation included localization and location-dependent metrics with a 20° threshold.The recordings contain synthesized events and spatial ambient noise from 15 indoor locations.Source: §4.1 · p. 3

4.2. Experimental settings

• Comparisons varied the representation and CRNN backbone; state-of-the-art runs used RD3Net with 1,024-frame inputs and rotation-based, vector-averaging post-processing.For the SELDnet two-branch baseline, the loss weight was 10; post-processing rotated FOA data, estimated and rotated back ACCDOA vectors, then averaged vectors across patterns.Source: §4.2 · pp. 3–4

4.3. Experimental results

• 3.4 points higher F20° and 1.0 point higher LRCD were reported than Two-branch for SELDnet in the testing split with CRNN.ACCDOA outperformed that representation for most metrics with both CRNNs and had fewer parameters.Source: Table 1 · p. 4, §4.3 · p. 4
• ACCDOA achieved the best LECD and F20° on the DCASE 2020 Task 3 development set with few parameters.The results also report that post-processing especially improves localization performance.Source: Table 2 · p. 4, §4.3 · p. 4
• The evaluation dataset allowed up to two overlapping classes, with almost all overlaps belonging to different classes; same-class overlap is identified as a case for a multi-track extension.Source: §4.3 · p. 4

• ACCDOA assigns sound-event activity to the length of its corresponding Cartesian DOA vector for SELD.Source: §5 · p. 4
• The representation enables single-target SELD with one network, avoiding objective balancing and increased network size.Source: §5 · p. 4
• On DCASE 2020 Task 3, ACCDOA outperformed two-branch representations in joint SELD metrics with fewer parameters and surpassed state-of-the-art systems on localization error and location-dependent F-score.Source: §5 · p. 4