Trimodal Disentangled Representation Learning for Motion Understanding
IEEE Access · 2026 · 14 · 84503-84516
DOI: 10.1109/access.2026.3699430 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Myungin Kim und 1 weiteren Autor:innen. Die Arbeit nennt 59 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Institute of Information & communications Technology Planning & Evaluation.
Abstract (Originalsprache)
Trimodal motion understanding systems that integrate motion, video, and text face a fund- amental challenge: existing approaches typically excel at either generation or retrieval, but rarely both. This trade-off arises from conflicting optimization objectives—generation requires modality-independent representations to preserve diversity, whereas retrieval benefits from tightly aligned cross-modal embeddings for accurate matching. To address this challenge, we introduce a TRImodAl representation Disentanglement framework (TRIAD) that decomposes trimodal embeddings into task-specific components: a generative subspace for generation and a discriminative subspace for retrieval. A shallow dual-router Mixture-of-Experts architecture enables independent optimization pathways for stable task-specific disentanglement. Our results demonstrate that explicit task-specific decomposition allows a single unified model to perform both generation and retrieval. TRIAD provides empirical analysis on jointly supporting generation and retrieval within a single trimodal framework.
Förderung
- Institute of Information & communications Technology Planning & Evaluation
Zitieren
APA
Kim, M., & Yu, R. (2026). Trimodal Disentangled Representation Learning for Motion Understanding. IEEE Access, 14, 84503-84516. https://doi.org/10.1109/access.2026.3699430
Vancouver
Kim M, Yu R. Trimodal Disentangled Representation Learning for Motion Understanding. IEEE Access. 2026;14:84503-84516. doi:10.1109/access.2026.3699430
BibTeX
@article{Kim2026,
author = {Kim, Myungin and Yu, Ri},
title = {Trimodal Disentangled Representation Learning for Motion Understanding},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {84503--84516},
doi = {10.1109/access.2026.3699430}
}
EUVIMED