Synthetic Depth Image-Based Category-Level Object Pose Estimation With Effective Pose Decoupling and Shape Optimization

Category-level object pose estimation is a crucial task in the field of computer vision and finds numerous applications. However, the presence of unknown objects, significant shape, and scale variations within the same category pose challenges in this task. To address these challenges and achieve efficient and accurate category-level object pose estimation, we present EffectPose in this article. We first observe that objects of the same category often possess similar key regions, such as handles on cups. These key regions can establish correspondences for spatial poses, enabling pose estimation. To facilitate this, we employ a segmentation network to divide point clouds into multiple parts and map them to a shared latent space. Subsequently, by considering the correspondences between predicted implicit models and real point clouds for various key regions, we accomplish pose estimation. Since real object point clouds are typically dense and contain outliers, we propose a novel point cloud sampling network that can accurately select representative points for efficient correspondence construction. Furthermore, we decouple the scale and pose of objects based on the SIM(3) invariant descriptor and propose an online pose optimization method using this descriptor. This method enables online prediction and optimization of poses. Finally, to enhance pose estimation accuracy, we introduce a distance-weighted pose optimization method for pose refinement and adjustment. Experimental results demonstrate that our proposed method achieves efficient pose estimation and generalization by utilizing only synthetic depth images and a minimal number of network parameters, surpassing the performance of most existing methods.