GENAI4E Team Achieves 90.92% mAP in AI City Challenge 2026 Track 4

The GENAI4E team has released a technical paper on arXiv (2608.12843) outlining their approach for the AI City Challenge 2026 Track 4, which centers on retrieving pedestrians exhibiting unusual behaviors through text-based descriptions. This task necessitates detailed reasoning about appearance, actions, interactions with objects, and contextual scenes. Their framework enhances a robust retrieval backbone and systematically incorporates diverse vision-language embedding models via score alignment and iterative ensemble fusion, concluding with disagreement-aware VLM reranking for ambiguous queries. Their method achieves impressive results on the official Pedestrian Anomaly Behavior (PAB) benchmark, with 90.92% mAP, 85.13% Recall@1, 97.72% Recall@5, and 98.68% Recall@10, showcasing the power of integrating complementary models. This cross-type submission highlights the complexity of cross-modal matching beyond standard text-based person retrieval.