RegRet: Enhancing Region-Level Retrieval in Large Multimodal Models
RegRet is presented, an LMM-based Region-level Retrieval framework that enhances the regional representations without compromising overall global retrieval performance and designs a multi-stage training pipeline that includes detailed localized captioning and regional contrastive learning tasks.
Xun Liang, Hong-Hui Yang, Weihang Pan et al.
· 0 citations