Seowoo Kim
Research Subjects
Seowoo Kim has a total of 4 co-authors. Here are the top co-authors by number of shared publications:
- Jeongin Ju1 shared publication
- Jungyu Choi
Department of Information and Communication Engineering, Soongsil University, Seoul, South Korea
1 shared publication Sungbin Im
1 shared publication- Sungbin Im
School of Electronic Engineering, Soongsil University, Seoul, South Korea
1 shared publication
Latest Publications
- Open Access
- 19 October 2026
- Published by Elsevier in SSRN Electronic Journal
Reference-Aware Branch-Consistent Prototype Transport for Domain-Incremental Acoustic Scene Classification
An acoustic scene recognizer must adapt to new recording environments whileretaining knowledge of earlier ones. In shared-label domain-incremental learning(DIL), stored class prototypes may become mismatched as domain-specific normalizationbranches define different representations and the shared backbonecontinues to change. We distinguish cross-domain discrepancy, measured acrossdomain–branch states at a fixed final checkpoint, from temporal drift, measuredfor a fixed domain–class state across checkpoints. We propose Branch-ConsistentPrototype Transport (BCPT), a parameter-free update that estimates classconditionedfeature displacement from current-domain data through the oldprototype’s retained branch. The update uses no earlier-domain training samplesand requires no test-domain identity. With CNN14 from Pretrained Audio NeuralNetworks (PANNs), BCPT reduces mean cosine distance to final-recomputedprototypes by 73.0% on TAU2019 and 45.9% on TAU2020, while increasing embeddingaccuracy by 0.67 and 1.29 percentage points over stage-stored prototypes.BCPT has the highest observed mean across four nonparametric estimators insix dataset–score comparisons, exceeding the next-best estimator by 0.08 to0.71 percentage points. A 2 × 2 estimator analysis yields positive interactionpoint estimates between same-class filtering and old-branch referencing (+0.21 to+1.54 percentage points), but the stage-branch control changes the branch andcheckpoint together and cannot isolate coordinate consistency. With EfficientATMN10,accuracy increases from 63.67% to 67.01%, compared with 67.41% forthe final-recomputed reference. Spectral comparisons show that cross-domaincompactness must be interpreted relative to between-class separation. Theseresults support explicitly defining prototype references and maintaining updatesin their stored branch coordinates.
- 19 October 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessReference-Aware Branch-Consistent Prototype Transport for Domain-Incremental Acoustic Scene Classification
An acoustic scene recognizer must adapt to new recording environments whileretaining knowledge of earlier ones. In shared-label domain-incremental learning(DIL), stored class prototypes may become mismatched as domain-specific normalizationbranches define different representations and the shared backbonecontinues to change. We distinguish cross-domain discrepancy, measured acrossdomain–branch states at a fixed final checkpoint, from temporal drift, measuredfor a fixed domain–class state across checkpoints. We propose Branch-ConsistentPrototype Transport (BCPT), a parameter-free update that estimates classconditionedfeature displacement from current-domain data through the oldprototype’s retained branch. The update uses no earlier-domain training samplesand requires no test-domain identity. With CNN14 from Pretrained Audio NeuralNetworks (PANNs), BCPT reduces mean cosine distance to final-recomputedprototypes by 73.0% on TAU2019 and 45.9% on TAU2020, while increasing embeddingaccuracy by 0.67 and 1.29 percentage points over stage-stored prototypes.BCPT has the highest observed mean across four nonparametric estimators insix dataset–score comparisons, exceeding the next-best estimator by 0.08 to0.71 percentage points. A 2 × 2 estimator analysis yields positive interactionpoint estimates between same-class filtering and old-branch referencing (+0.21 to+1.54 percentage points), but the stage-branch control changes the branch andcheckpoint together and cannot isolate coordinate consistency. With EfficientATMN10,accuracy increases from 63.67% to 67.01%, compared with 67.41% forthe final-recomputed reference. Spectral comparisons show that cross-domaincompactness must be interpreted relative to between-class separation. Theseresults support explicitly defining prototype references and maintaining updatesin their stored branch coordinates.
- Open Access
- 28 July 2026
- Published by Elsevier in Results in Engineering
- Vol. 32
Multi-task U-Net with class-conditional Mahalanobis scoring for drone acoustic anomaly detection
- 28 July 2026
- Published by Elsevier in Results in Engineering
- Vol. 32
Open AccessMulti-task U-Net with class-conditional Mahalanobis scoring for drone acoustic anomaly detection