Reproduce “Heterogeneous Vision-Language Ensemble with Disagreement-Aware Reranking for Text-Based Person Anomaly Retrieval” on OpenResearch

About this paper

Huu-An Vu, Cam Tu Tran Thi, Thanh Toan Le Ngo, Hoang Vo, Do Trung Hieu, Hieu Dinh Trung Pham, Khang Minh Le, Huy Minh Nhat Nguyen

Researchers developed a heterogeneous vision-language ensemble with a disagreement-aware reranking mechanism for text-based person anomaly retrieval, achieving 90.92% mAP on the AI City Challenge 2026 Track 4 benchmark. This system effectively combines diverse vision-language models and selectively applies a computationally expensive VLM reranker to ambiguous queries, improving accuracy while maintaining efficiency.

Key insights

• Integrating a diverse set of vision-language embedding models trained with different architectures and objectives significantly enhances the robustness and accuracy of text-based person anomaly retrieval.
• Careful weight tuning during iterative ensemble fusion is critical, especially for large vision-language models, to maximize performance gains on challenging datasets.
• An adaptive reranking strategy, which employs computationally expensive VLMs only for ambiguous queries identified by model disagreement, effectively balances high accuracy with practical inference efficiency.

Summary by alphaXiv.