6D pose estimation of category-level objects based on keypoint enhancement

Category-level object 6D pose estimation aims to predict 6D pose and size data for unseen target instances with known categories, requiring models to possess strong intra-category generalization capabilities for objects of diverse shapes. Although many previous methods have attempted to enhance the learning of features with significant intra-category variations, they overly emphasize point-wise dense feature learning or point-wise coordinate prediction while lacking the learning of target-wide geometric features and structures. This results in poor generalization performance for unseen target instances with significant shape variations and occluded scenarios in pose estimation. To address this issue, we propose a category-level object 6D pose estimation method based on keypoint enhancement. First, this method constructs an instance-adaptive keypoint detection module through a dual-modal separated cross-modal attention mechanism and adaptive weight fusion, learning a set of sparse keypoints to represent the overall schematic structural information of the instance. Second, it builds a perceptual feature aggregation module via adaptive dynamic neighborhood construction, linear and sinusoidal positional encoding, and multi-head attention mechanisms, integrating global features through conditional normalization to adaptively learning local and global features of the current target instance’s keypoints. The keypoint learning module can fully consider and exploit information from different modalities to effectively learning the key structure and schematic information of unseen instances. The feature aggregation module dynamically constructs local and global geometric information of the keypoints based on the keypoint foundation, thereby learning geometric information of unseen instances and improving the model’s ability to learn geometric and structural information of unseen instances, as well as its generalization capability for pose estimation of unseen target instances. Our extensive experimental results on the REAL275 and CAMERA25 datasets demonstrate that the proposed keypoint enhancement method has excellent performance.