Publikation
Hierarchy-Aware Representation Learning for Bio-acoustic Data
Keyhyun Ku; Rida Saghir; Thiago Gouvea; Daniel Sonntag
In: Computer Vision for Ecology. Computer Vision for Ecology (CV4E-2026), 3rd Workshop on Computer Vision for Ecology at the 19th European Conference on Computer Vision, located at ECCV-2026, September 8-12, Malmö, Sweden, Springer, 2026.
Zusammenfassung
Bio-acoustic data is rich in information and carries meaningful hierarchies that can support different objectives towards ecological conservation. We take hierarchy-aware training objectives, originally developed for computer vision applications and transfer them to bioacoustic audio classification. We adapt four hierarchical learning objectives to audio embeddings from frozen domain-specific and multimodal audio models, and evaluate them on both a biological taxonomic hierarchy and a soundscape hierarchy. Across embedding models, hierarchy types, and single- vs. multi-label settings, we find that imposing hierarchical constraints improves classification even at finest and coarser level. While no single method wins universally, we observe larger margins in the multi-label PAM setting, and Focal Tree-Min performs best most often.