基于类别级先验重建的六维位姿估计方法

6D pose estimation based on category-level prior reconstruction

针对现有物体六维姿态估计方法难以泛化到真实场景、未知新物体的问题,文章提出了一种基于类别级先验重建的物体六维位姿估计方法,通过提取并融合输入物体与先验模板的三维特征,将先验模板由初始的局部坐标系变换到当前相机坐标系下,根据变换前后的点对匹配关系求得目标物体的六维位姿,通过这种方式能够使得模型更关注类别级共有信息、抑制对输入物体的特异性几何结构表征,对未知新物体具有更好的泛化能力。实验表明,当训练集仅包含合成数据时,该方法在真实数据上具有更好的性能。

Existing methods always lack ability to generalize to real scenes and unknown novel objects,This paper proposes a method for 6D object pose estimation based on category-level prior reconstruction.By extracting and fusing 3D features of the input object and the prior template,the prior template is transformed from the initial local coordinate system to the current camera coordinate system,The pose of the target object is obtained according to the point pair matching before and after this transformation.In this way,the trained model naturally pays more attention to category-level shared information,while less to distinctive feature from the input novel object.Experiments show that this method has better performance on real data when the training set contains only synthetic data.