Enhancing Category-Level 6D Object Pose Estimation with a Point Cloud Completion and Deformation Network

Category-level pose estimation aims to estimate the rotation, translation and size of unseen instances in specified categories. This task becomes challenging due to intra-class variation and the intricate nature of estimating dense correspondences between instance cloud and NOCS point cloud. In this paper, we propose a patch-seeds-based point cloud completion and deformation network to address these challenges. Our method leverages patch seeds to complete instance point cloud and deform categorical shape prior, enabling us to estimate complete instance point cloud in both camera space and normalized space. Subsequently, we employ the ICP algorithm to directly register the estimated point cloud and derive the pose. This method not only robustly cope with intra-class variations, but also avoids the complex and dense correspondences between instance point cloud and NOCS point cloud by directly estimating the network. Our designed network adopts a multi-stage point cloud completion and deformation strategy, where each stage progressively refines the accuracy of the point cloud reconstruction. We conducted extensive experiments on a recognized category-level 6D attitude estimation benchmark, and the experimental results showed that our proposed method significantly improved the accuracy of attitude estimation compared to previous methods.