Action-Aligned Retrieval with Pairwise Multimodal Reranking for Text-Based Person Anomaly Search
This work proposes a unified three-stage coarse-to-fine framework that combines action-aligned retrieval with pairwise multimodal reranking to bridge the pose-semantic gap, and proposes an efficient off-the-shelf reranking module that leverages a pivot-promote algorithm to perform direct pairwise visual comparisons.