慧眼识"新"开放词汇目标检测OA

To address the low detection accuracy of novel classes in open-world scenarios-primarily caused by weak foreground discrimination and strong bias toward base classes-an open-vocabulary object detection framework named Sharp Eyes Spot the"Novel"in Open-Vocabulary Object Detection(SSN-OVD)is proposed.First,a Foreground Feature Discrimination(FFD)module is introduced,in which a fore-ground estimator is employed to model potential novel-class regions and generate high-quality pseudo-la-bels,enabling more precise foreground-background separation and enhancing the discriminability of fore-ground features.Second,a Bidirectional Feature Alignment(BFA)module is designed to leverage bidi-rectional cross-modal alignment together with confidence calibration,thereby mitigating base-class bias during training and strengthening the model's capability to learn robust representations of novel classes.Third,a Contrastive Denoising Training(CDT)module is developed,incorporating noisy visual-text pairs into the contrastive learning process to further improve feature discrimination and generalization for novel categories.Experimental results demonstrate that the proposed approach achieves state-of-the-art performance,yielding novel-class detection accuracies of 44.9% on COCO and 37.4% on the more chal-lenging fine-grained LVIS dataset.These results indicate that the method effectively enhances novel-class detection in open-world environments.