Open World Object Detection (OWOD) aims not only to recognize known classes but also to discover unknown objects and incrementally learn them in subsequent stages. In OWOD, complex background clutter hinders the accurate recognition of unknown objects, while the availability of only known-class labels further biases model predictions toward known categories. To address these issues, we propose a novel OWOD framework termed SGAC-OWOD. It consists of two key components: a Saliency-Guided Foreground Enhancement (SFE) module and an Adversarial Perturbation-based Classification Calibration (APCC) module. In SFE, we introduces a class-agnostic saliency prior to enhance foreground representations through adaptive multi-scale feature fusion and box-level filtering. In APCC, we applies multi-step adversarial perturbations to semantically enhance proposal features and calibrate classification confidence according to their stability under perturbations, effectively mitigating label bias. We conduct comprehensive evaluations on MS-COCO and PASCAL VOC. On Task 1, the unknown recall (U-Recall) reaches 43.7%, improving by about 7.6% over the state-of-the-art method. In addition, we conducted cross-dataset experiments on other benchmarks. Our method achieves the best performance on most key metrics, further demonstrating its cross-dataset generalization ability.
Wang et al. (Fri,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: