Scholarlink

<p>Unknown objects in open-world scenarios pose significant challenges to object detection accuracy. Existing open-set detection methods often rely on manual annotation of unknown objects followed by incremental learning, which is both time-consuming and labor-intensive. To discover the potential categories of unknown objects and reduce the cost of manual annotation, we propose OWDet, an open-world object detection framework designed to automatically discover and learn new object concepts. A key component of OWDet is a semi-supervised contrastive clustering with strong-weak augmentation algorithm that effectively organizes unknown objects into distinct clusters. By leveraging labels from known objects, the algorithm guides the clustering process, facilitating the discovery of novel concepts that may not exist in the original training set. To ensure reliability, a cluster filtering mechanism is introduced to generate pseudo-labels only for high-confidence clusters, thereby reducing noise and improving learning efficiency. Once novel classes are identified, they are combined with known categories to retrain the detector, enhancing its overall detection performance. Extensive experiments on a composite dataset combining KITTI, Pascal VOC, and MS COCO demonstrate that OWDet significantly outperforms existing methods in open-world scenarios. Notably, it achieves a 94.7% purity in new category discovery across 942 unknown samples. Furthermore, OWDet maintains high detection recall on cross-domain tests without fine-tuning, confirming its robustness in open-world environments. Our source code is available at <a href="https://github.com/StarkM10/OWDet">https://github.com/StarkM10/OWDet</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Semi-supervised contrastive clustering with strong-weak augmentation for novel class discovery in open-world object detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月4日

查看全文
  • Fei Wang,
  • Rui Ma

摘要

Unknown objects in open-world scenarios pose significant challenges to object detection accuracy. Existing open-set detection methods often rely on manual annotation of unknown objects followed by incremental learning, which is both time-consuming and labor-intensive. To discover the potential categories of unknown objects and reduce the cost of manual annotation, we propose OWDet, an open-world object detection framework designed to automatically discover and learn new object concepts. A key component of OWDet is a semi-supervised contrastive clustering with strong-weak augmentation algorithm that effectively organizes unknown objects into distinct clusters. By leveraging labels from known objects, the algorithm guides the clustering process, facilitating the discovery of novel concepts that may not exist in the original training set. To ensure reliability, a cluster filtering mechanism is introduced to generate pseudo-labels only for high-confidence clusters, thereby reducing noise and improving learning efficiency. Once novel classes are identified, they are combined with known categories to retrain the detector, enhancing its overall detection performance. Extensive experiments on a composite dataset combining KITTI, Pascal VOC, and MS COCO demonstrate that OWDet significantly outperforms existing methods in open-world scenarios. Notably, it achieves a 94.7% purity in new category discovery across 942 unknown samples. Furthermore, OWDet maintains high detection recall on cross-domain tests without fine-tuning, confirming its robustness in open-world environments. Our source code is available at https://github.com/StarkM10/OWDet.