System view / research question
Pose → embedding → retrievalResearch pipeline from corrupted pose sequence to cosine motion retrievalA noisy pose query passes through frozen MotionBERT and a trained linear head into an embedding space, then is ranked against a clean motion gallery using cosine similarity. Contextual neighborhoods shape training only. This is an explanatory animation, not a measured result.
Robust Human-Motion Retrieval from Pose Sequences
Funded undergraduate research comparing motion-retrieval training recipes under joint jitter, missing joints, and missing frames.
Read study
