Heterogeneous Vision-Language Ensemble with Disagreement-Aware Reranking for Text-Based Person Anomaly Retrieval

The GENAI4E framework utilizes heterogeneous vision-language ensemble fusion and score alignment to improve text-based person anomaly retrieval. By applying disagreement-aware VLM reranking to ambiguous queries, the method achieves 90.92% mAP and 85.13% Recall@1 on the Pedestrian Anomaly Behavior benchmark.