MoCHA framework improves motion-text retrieval by filtering captions to focus on motion-specific content rather than annotator bias

Text-motion retrieval systems struggle because different annotators describe the same motion differently, mixing recoverable motion semantics with personal style and context

Standard contrastive training treats each caption as a single positive example, ignoring this distributional structure and creating embedding variance that weakens alignment

MoCHA addresses this by canonicalizing captions to extract only motion-recoverable content (action type, body parts, directionality) before encoding

The canonicalization approach produces tighter positive clusters and better-separated embeddings across the shared embedding space