Detect in Any Scene: An Agentic Framework for Object Detection with Experience-Aware Reasoning
Researchers have introduced DetAS, an agentic framework for object detection that treats the task as a dynamic decision process. This framework utilizes a Multimodal Large Language Model (MLLM) to adaptively compose detection workflows by selecting from a toolbox of restoration modules and specialized detectors. The extended DetAS-X version further refines decision quality by accumulating experience from annotated data, enabling it to progressively adapt its policy during inference. Experiments show DetAS-X significantly outperforms existing MLLM-based detectors, achieving substantial gains in F1 score on challenging benchmarks. AI
IMPACT Introduces a novel agentic approach to object detection, potentially improving performance in complex, dynamic environments.