Symmetry-Aware 9D Pose Estimation with Sim(3)-Consistent Feature and Spherical Inception Convolution
Researchers have developed a novel method for category-level object pose estimation, addressing limitations in generalization to unseen objects. The approach utilizes a semantic-guided symmetry-aware module, leveraging a large vision model (LVM) to accurately infer translation and size without requiring shape priors. Additionally, a feature fusion module combines LVM semantic features with geometric features using a spherical inception convolution to model long-range dependencies efficiently. This method achieves state-of-the-art results on benchmarks and has been applied to develop a robotic picking system. AI
IMPACT Enhances robotic perception and manipulation capabilities by improving object pose estimation accuracy and generalization.