ClearPose

使用RealSense L515相机在室内环境中捕获63个透明对象的ClearPose数据集。它包含RGB，原始深度，地面真实深度 (由渲染对象的深度超过原始深度重叠生成)，地面真实表面法线 (根据地面真实深度计算) 图像以及所有对象实例6D姿势。数据注释是通过工具progresslaabeler完成的，该工具包括解决相机轨迹的ORB-SLAM3，用于对象姿势的Blender界面以及用于生成数据的后端渲染器。对象的几何模型是在Blender中手动创建的，并在注释过程中进行验证。我们还包括一些场景中出现的不透明对象 (来自YCB和HOPE数据集) 的模型。

The ClearPose dataset captures 63 transparent objects in indoor environments using a RealSense L515 camera. It contains RGB images, raw depth maps, ground-truth depth maps (generated by overlaying rendered object depths onto the original depth data), ground-truth surface normal images (calculated from the ground-truth depth maps), as well as 6D poses for all object instances. The data annotations were completed using the ProgressLabeler tool, which integrates ORB-SLAM3 for camera trajectory optimization, a Blender interface for object pose annotation, and a backend renderer for dataset generation. The geometric models of the objects were manually created in Blender and validated during the annotation procedure. We also include models of some opaque objects sourced from the YCB and HOPE datasets that appear in the scenes.

搜集汇总

背景与挑战

背景概述

ClearPose是一个由密歇根大学发布的透明物体识别数据集，包含RGB、深度、表面法线图像及6D姿势信息，适用于计算机视觉和机器人抓取研究。