MIVPG on E-commerce: Multi-Image/Multi-Patch Aggregation for Captioning

MIVPG uses hierarchical MIL to outperform patch concatenation and single-image baselines, proving CSA is key for correlation.