A Generalized Open-World Referring Expression Segmentation Approach to Biometric Forensics with Generative Multi-Agent System

In open-world biometric recognition, ambiguous descriptions, missing discriminative features, large-scale image collections, and potential target absence pose significant challenges. To address these issues, we propose Generative Multi-Agent Referring Expression Segmentation (GMA-RES), a method that localizes and segments biometric targets from natural language descriptions for biometric applications such as surveillance and forensic search. GMA-RES formulates open-world biometric recognition as an autonomously orchestrated multi-agent decision process, where a vision-language Planner Agent performs high-level semantic reasoning and dynamically coordinates multiple domain-specific Specialist Agents for image retrieval, localization, pixel-level segmentation, and verification. This modular design decouples semantic reasoning from biometric evidence generation, enabling adaptive multi-image reasoning and explicit absence handling without modifying pre-trained model architectures. A two-stage training paradigm further optimizes both reasoning accuracy and inference efficiency. Experimental evaluation on open-world biometric recognition tasks demonstrates that GMA-RES achieves an accuracy of 57.8%, significantly outperforming two-stage visual prompting methods (30.5%) and one-stage generative vision-language models (50.2%). The results validate the superior performance of GMA-RES in handling ambiguous queries, multi-image reasoning, and absent-target scenarios, providing a scalable and interpretable solution for open-world biometric applications.