New CG-CLIP framework uses AI-generated captions to identify people in challenging video scenarios like sports and dance performances.

CG-CLIP combines CLIP model with multi-modal large language models to generate detailed textual descriptions for person re-identification

Caption-guided Memory Refinement (CMR) component captures fine-grained identity-specific features using captions from MLLMs

Token-based Feature Extraction (TFE) uses cross-attention with learnable tokens to efficiently aggregate spatiotemporal features and reduce computational overhead

Method addresses high-difficulty scenarios where multiple individuals wear similar clothing and perform dynamic movements, such as in sports and dance