Voxel51/IndustryShapes

IndustryShapes是一个用于工业装配工具和组件6D位姿估计的RGB-D基准数据集。它提供了五个具有挑战性的工业对象的高质量标注数据，这些对象具有弱纹理、反射表面、对称性和薄结构等特点，并在真实的工业装配环境中捕获。数据集旨在支持实例级和新对象位姿估计方法，并且是首个包含用于无模型方法的RGB-D静态上机序列的数据集。数据集包含13012个样本，涵盖RGB图像、深度图、相机内参、位姿注释（如旋转矩阵、平移向量）、边界框和实例分割掩码等字段。数据采集使用Intel RealSense传感器，并遵循BOP规范。注释过程包括基于标记、SfM半自动和合成方法，平均深度误差小于12毫米。数据集适用于6D位姿估计、对象检测、实例分割、深度估计研究和机器人操作等任务。

IndustryShapes is an RGB-D benchmark dataset for 6D object pose estimation of industrial assembly tools and components. It provides high-quality annotated data of five challenging industrial objects—characterized by weak texture, reflective surfaces, symmetries, and thin structures—captured in realistic industrial assembly environments. The dataset is designed to support both instance-level and novel-object pose estimation approaches, and is the first dataset to include RGB-D static onboarding sequences for model-free methods. The dataset contains 13012 samples, including RGB images, depth maps, camera intrinsics, pose annotations (such as rotation matrices and translation vectors), bounding boxes, and instance segmentation masks. Data were captured using Intel RealSense sensors and formatted according to the BOP specification. Annotation processes involved marker-based, SfM semi-automatic, and synthetic methods, with mean absolute depth error less than 12 mm. The dataset is suitable for tasks like 6D pose estimation, object detection, instance segmentation, depth estimation research, and robotic manipulation.