EUVIMEDEUVIMED entdecken
Crossref

VMR: Training‐Free 3D Motion Retrieval From Video

Jaeho Im, Minhye Kang

Computer Animation and Virtual Worlds · 2026-07 · 37 (4)

DOI: 10.1002/cav.70166 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Computer Animation and Virtual Worlds, online erschienen am 10.08.2026. Verfasst von Jaeho Im und 1 weiteren Autor:innen. Die Arbeit nennt 29 Literaturangaben. Gefördert von: Korea Creative Content Agency.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

We present Video Motion Retriever (VMR), a framework that retrieves 3D human motions from a database using video as the input query. Given a reference video, VMR extracts 3D body motion via segmentation and whole‐body pose estimation, encodes it into a latent vector using a pretrained motion encoder, and retrieves the most similar motion through cosine similarity matching. Since both queries and database entries share the same encoding pipeline, new motions can be registered without retraining. We evaluate VMR on a commercial motion capture database of 1327 motions through multiple protocols. In Self‐Retrieval, VMR achieves 97.81% Top‐1 Accuracy. In a direct comparison against TMR on the Eyes_Japan dataset, VMR reaches 99.66% Top‐1 Accuracy compared to 77.48% for the strongest text‐based baseline. In Cross‐Domain Retrieval with out‐of‐domain videos, VMR retrieves kinematically faithful motions with a mean top‐1 cosine similarity of 0.899 across 18 queries, despite substantial visual domain gaps. In Open‐Domain Retrieval without ground‐truth matches, VMR records a mean cosine similarity of 0.927. Ablation studies further confirm the contribution of the segmentation and hand reconstruction modules. These results demonstrate that VMR provides a practical, training‐free, and rig‐agnostic solution for video‐based motion retrieval.

Förderung

  • Korea Creative Content Agency

Zitieren

APA

Im, J., & Kang, M. (2026). VMR: Training‐Free 3D Motion Retrieval From Video. Computer Animation and Virtual Worlds, 37(4). https://doi.org/10.1002/cav.70166

Vancouver

Im J, Kang M. VMR: Training‐Free 3D Motion Retrieval From Video. Computer Animation and Virtual Worlds. 2026;37(4). doi:10.1002/cav.70166

BibTeX
@article{Im2026,
  author = {Im, Jaeho and Kang, Minhye},
  title = {VMR: Training‐Free 3D Motion Retrieval From Video},
  journal = {Computer Animation and Virtual Worlds},
  year = {2026},
  volume = {37},
  number = {4},
  doi = {10.1002/cav.70166}
}
Kostenlos starten →