Adaptive objectness learning for enhanced unknown object detection

Unknown object detection requires training models on labeled known objects to generalize to unknown categories. A key challenge lies in effectively learning objectness from known objects to represent unknowns. Two issues limit the detection performance of existing methods: (1) non-adaptive feature extraction and fusion, which causes some unknown objects with diverse semantics and scales to be overlooked; and (2) inflexible objectness learning, which hinders accurate localization of unknown objects. To address these issues, we propose the adaptive unknown object detector (AUD), which enhances objectness learning for improved unknown object detection. First, AUD introduces the multi-scale feature adaptive fusion module (MFAFM) to adaptively fuse cross-scale features, providing rich semantic information for subsequent objectness learning. Second, the adaptive objectness score (AOS) is proposed to learn generalized objectness knowledge from the positional relationships of known objects, enabling accurate localization of unknown object boundaries. During inference, AOS takes fused features from MFAFM as input and outputs a set of potential objects. Finally, a boxes adaptive determination strategy is designed to filter redundant potential objects and retain accurate prediction results. Experimental results demonstrate that AUD significantly outperforms state-of-the-art methods, achieving 18.3% and 11.4% absolute gains in unknown precision rate on the COCO-OOD and COCO-Mixed benchmarks, respectively. Our code and datasets are publicly available to facilitate research reproducibility and future advancements in this field. Our code is publicly available at https://github.com/ndwxhmzz/AUD.