Fusing shape descriptors and geometric details for robust category-level object pose estimation

Accurate estimation of object poses and sizes within specific categories is crucial for applications such as robotic manipulation and scene understanding. Despite recent advances, significant intra-category shape variations pose a great challenge, reducing the accuracy and robustness of shape prior-based methods. This paper proposes a novel network that leverages shape descriptors and local geometric features for category-level object pose estimation. By capturing geometric structures of the object through shape descriptors, our approach effectively handles shape variations and efficiently distinguishes between instances within the same category. Additionally, we design a local feature detector to extract fine-grained geometric details for enhancing shape descriptor-guided learning. Moreover, an attention mechanism is employed to adaptively highlight significant features, improving the model’s robustness for objects with complex structures. Our network also includes a confidence-based pose estimator that assigns a confidence score to each pose prediction. This integration allows for the acquisition of accurate poses with high confidence by penalizing poor poses with low confidence. Experimental results on the CAMERA25 and REAL275 datasets demonstrate the effectiveness of the proposed network, which achieves accuracy improvements of 5.1 and 12.5, respectively, under the 52cm metric compared to state-of-the-art methods. These results underscore our network’s superiority in handling objects with large shape variations and complex structures. The code will be released at https://github.com/yliu1999/Shape-Descriptor.