Researchers have developed a novel framework for text-based person anomaly search, addressing the challenge of transferring models trained on synthetic data to real-world surveillance images. The system employs a multi-model retrieval and reranking approach, adapting the Qwen3-VL-Embedding-8B model with techniques like bidirectional contrastive learning and domain randomization. It fuses scores from multiple Qwen3-VL checkpoints and a SigLIP2 model, then uses a Qwen3-VL-Reranker-8B cross-encoder for final evaluation. This method achieved strong results, including 94.7815% mAP and 90.8493% R@1 on the AI City Challenge 2026 Track 4, securing an eighth-place finish on the public leaderboard. AI
IMPACT This research advances synthetic-to-real transfer for vision-language models, potentially improving surveillance and security applications.
RANK_REASON This is a research paper detailing a novel AI framework for a specific computer vision task. [lever_c_demoted from research: ic=1 ai=1.0]
- AI City Challenge 2026 Track 4
- arXiv
- Hugging Face
- Qwen3-VL-Embedding-8B
- Qwen3-VL-Reranker-8B
- SigLIP2
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
