Semantic-Degrade Learning Framework for Open World Object Detection

Open World Object Detection (OWOD) challenges the conventional closed-world assumption of traditional object detection models, addressing the dynamic nature of real-world scenarios where systems encounter unknown objects. Unlike existing OWOD approaches which often rely on manually selected unknown proposals, we introduce an Adaptive Semantic-Degrade Learning framework. This framework, inspired by cognitive development theory, guides the model to capture more low-level semantic features by naturally degrading the learned feature patterns, thereby enabling the discovery of unknown targets that share these features with different known classes. We implement this framework based on the closed-set object detector Deformable DETR, constructing a reusable open world architecture of two branches for known class detection and unknown object detection. Extensive experiments on the common-used benchmark validate the progressiveness of our framework. The experimental results show that compared with other state-of-the-art methods, our model achieves nearly 50% improvement in unknown mAP and even higher known detection performance, demonstrating excellent detection performance.