基于深度三维模型表征的类别级六维位姿估计

3D deep implicit function for category-level object 6D pose estimation

类别级物体六维位姿估计在机器人操作、自动驾驶和增强现实等领域有着广泛的应用。相较于实例级任务,类别级六维位姿估计的难点主要在于类别先验特征基础上的类内差异。本文采用一种基于有向距离场(Signed Distance Field,SDF)的深度三维模型表征提取出类别级先验共享信息,同时依据输入深度图像的几何形状特征搜索最优的形状隐变量,两者结合重建出标准空间内的完整实例模型。通过学习深度点与标准化实例模型的点对匹配关系,即可求解出物体的六维位姿参数。实验证明本文提出的类别级六维位姿估计架构具有良好的性能和对类内新物体的泛化能力。

Category-level object 6D pose estimation is important for the task of robot manipulation,autonomous driving and augmented reality.Compared with instance-level one,the challenge of category-level 6D pose estimation mainly lies in the intra-class variation given a category prior.In this paper,a deep implicit function for representing 3D model based on SDF is adopted to extract the shared category-level prior.At the same time,the optimal shape latent code is predicted according to the geometric feature extracted from the input depth image.Both of the shared prior decoder and the specific shape latent code are combined together to reconstruct the complete instance in the normalized canonical space.Then the 6D pose could be solved by estimating the point matching between the depth point cloud and the canonical instance.Experiments show that the proposed framework for category-level 6D pose estimation achieves relatively good performance as well as generalization ability for novel instances within the same category.