Zoo3D：场景级零样本 3D 目标检测
Zoo3D: Zero-Shot 3D Object Detection at Scene Level
提出无需训练的零样本 3D 目标检测框架，利用 2D 实例掩码的图聚类构建 3D 包围盒，并通过最佳视角选择与视角一致性掩码生成进行开放词表语义标注，可直接用于点云或图像场景的 3D 感知。
原文依据 We take this further with Zoo3D, the first training-free 3D object detection framework.
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论3D 目标检测是空间理解的基础。真实环境要求模型能够识别多样且先前未见的物体，而这仍是闭集方法的主要局限。现有的开放词表 3D 检测器放宽了标注要求，但仍依赖训练场景的点云或图像。我们更进一步，提出了 Zoo3D，这是首个无需训练的 3D 目标检测框架。我们的方法通过 2D 实例掩码的图聚类构建 3D 包围盒，并使用一种新颖的开放词表模块分配语义标签，该模块包含最佳视角选择和视角一致性掩码生成。Zoo3D 以两种模式运行：零样本 Zoo3D_0 无需任何训练，自监督 Zoo3D_1 通过在 Zoo3D_0 生成的伪标签上训练一个类别无关的检测器来优化 3D 包围盒预测。我们进一步将 Zoo3D 扩展到点云之外，可直接在带位姿甚至无位姿的图像上运行。在 ScanNet200 和 ARKitScenes 上，Zoo3D_0 和 Zoo3D_1 均在开放词表 3D 目标检测中取得了最优结果，其中零样本 Zoo3D_0 优于所有现有的自监督方法。
英文原文摘要 1165 字符
3D object detection is fundamental for spatial understanding. Real-world environments demand models capable of recognizing diverse, previously unseen objects, which remains a major limitation of closed-set methods. Existing open-vocabulary 3D detectors relax annotation requirements but still depend on training scenes as point clouds or images. We take this further with Zoo3D, the first training-free 3D object detection framework. Our method constructs 3D bounding boxes via graph clustering of 2D instance masks and assigns semantic labels using a novel open-vocabulary module with best-view selection and view-consensus mask generation. Zoo3D operates in two modes: the zero-shot Zoo3D_0, which requires no training, and the self-supervised Zoo3D_1, which refines 3D box prediction by training a class-agnostic detector on Zoo3D_0-generated pseudo labels. We further extend Zoo3D beyond point clouds to operate directly on posed or even unposed images. Across ScanNet200 and ARKitScenes, both Zoo3D_0 and Zoo3D_1 achieve state-of-the-art results in open-vocabulary 3D object detection, with zero-shot Zoo3D_0 outperforming all existing self-supervised methods.