CPPN: Cross Perception Point Net for Pose Estimation

Keypoint recognition method based on dense correspondence represents a cutting-edge approach to 6-degree-of-freedom (6DoF) pose estimation. By fully leveraging pixel-level information, this method effectively addresses the challenges posed by occlusion and truncation in pose estimation. However, it exhibits limited robustness in cluttered scenes or scenarios involving multiple objects. Moreover, network complexity often lead to difficulties in achieving convergence during training. To address these challenges, this paper proposes a pose estimation network, CPPN, based on a multi-task learning architecture. The network employs image segmentation and center offset prediction tasks to assist keypoint prediction. Additionally, feature tensors are introduced to enhance the effects of transfer learning among the three tasks. A gradient normalization algorithm is also incorporated, enabling the network to adaptively balance the weights among the three tasks, ensuring similar learning rates and preventing overfitting. The trained model was evaluated on the Linemod dataset, and the results demonstrate that the CPPN model outperforms existing state-of-the-art multi-task learning-based pose estimation models. It effectively resolves the robustness issues of dense correspondence methods in cluttered scenes and addresses the convergence challenges in 3D keypoint prediction tasks.