CNN-based 3D Pose Estimation for Weakly Textured Rotationally Symmetric Targets

Abstract

Current 3D pose estimation methods face challenges in handling rotationally symmetric objects and generalizing to unseen objects, which limits their real-world applicability. To address these issues, this paper proposes a novel 3D pose estimation method. First, the distribution of sampled viewpoints is designed based on the potential symmetry orders of the target object. Second, a Convolutional Neural Network (CNN) is employed to match the sampled images, predicting the symmetry order of the rotationally symmetric target. This predicted order is then used to map the poses of the sample images to a canonical rotation space, effectively eliminating rotational ambiguity. Finally, the target image is matched against the sample images using a CNN to estimate its 3D pose based on the matching results. Compared with existing template-matching and regression-based approaches, our method not only resolves the training ambiguity caused by rotational symmetry but also achieves competitive generalization to unseen objects, with a simpler network architecture and lower computational overhead. Comparative experiments demonstrate that the proposed method resolves the issue of directly training on rotationally symmetric objects with CNNs and enables symmetry and pose estimation for unseen objects, thereby enhancing the practicality of 3D pose estimation in real-world scenarios.