Sound Event Classification on ESC-50 Simulated Distributed Layouts (five-fold cross-validation)
36.8Accuracy (Circular)Oracle (beamform)
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Oracle (beamform)Backbone=AST, Processing=delay-and-sum beamforming, Ground truth=sound source location, Evaluation protocol=Averaged over training layouts2026.01 | 36.8 | 38.6 | 36.9 | 37.4 | |
| Oracle (MaxSNR)Backbone=AST, Processing=Maximum-SNR channel selection, Evaluation protocol=Averaged over training layouts2026.01 | 33.4 | 36.9 | 33.9 | 34.7 | |
| ProposedBackbone=AST, Inpainting=RTM-based, Evaluation protocol=Averaged over training layouts2026.01 | 22.3 | 25.8 | 22.8 | 23.6 | |
| Ch swapBackbone=AST, Augmentation=channel swapping, Evaluation protocol=Averaged over training layouts2026.01 | 22.1 | 19 | 10.9 | 17.3 | |
| Ch selectorBackbone=AST, Mechanism=scaling sparsemax based channel selection, Evaluation protocol=Averaged over training layouts2026.01 | 20 | 19.9 | 14.4 | 18.1 | |
| BaselineBackbone=AST, Evaluation protocol=Averaged over training layouts2026.01 | 17.4 | 16.7 | 9.7 | 14.6 |