Symmetry-Aware 9D Pose Estimation with Sim(3)-Consistent Feature and Spherical Inception Convolution

Researchers proposed a symmetry-aware category-level 9D pose estimation method for unseen objects.

It combines a translation-and-size estimator guided by a large vision model with a spherical large-kernel inception convolution module.

The approach better handles intra-class variation and the non-linear Sim(3) pose space, improving rotation and pose estimation.

It reports state-of-the-art results on benchmarks and real-world scenes and has been applied in a robotic picking system.