HOPE

HOPE数据集，由NVIDIA创建，专注于家庭环境中物体的6自由度姿态估计，特别适用于机器人操作研究。该数据集包含28种玩具杂货物品，这些物品易于购买且适合机器人抓取和操作。数据集提供了这些物品的3D扫描纹理模型，适合生成合成训练数据，以及在具有部分遮挡、极端光照变化、多图像实例和大量姿态变化的杂乱场景中的RGBD图像。通过半自动RGBD到模型纹理对应关系，图像被标注了精确到几毫米的地面实况姿态。此外，数据集还提出了基于匈牙利分配算法的新姿态评估指标ADD-H，该指标对物体几何中的对称性具有鲁棒性，无需明确枚举。

The HOPE dataset, created by NVIDIA, focuses on 6-degree-of-freedom (6DoF) object pose estimation in household environments and is specifically tailored for robotic manipulation research. This dataset includes 28 types of toy grocery items, which are readily purchasable and suitable for robotic grasping and manipulation. It provides 3D scanned textured models of these items for generating synthetic training data, as well as RGBD images captured in cluttered scenes with partial occlusion, extreme lighting variations, multiple image instances and significant pose variations. Through semi-automatic RGBD-to-model texture correspondence, the images are annotated with ground-truth poses accurate to within a few millimeters. Additionally, the dataset proposes a novel pose evaluation metric ADD-H based on the Hungarian assignment algorithm, which is robust to symmetries in object geometry without requiring explicit enumeration.