GSV-Pose: Pose estimation based on geometric similarity voting

Object pose estimation is a fundamental problem in 3D computer vision and has gained significant attention with the rapid advancements in autonomous driving, robotics, and augmented reality. Traditional voting-based approaches often suffer from reduced accuracy when dealing with partially observed objects. To overcome this limitation, our method incorporates a superpoint matching network to compute local geometric similarities, which effectively guides the voting process and enhances pose estimation robustness. Experimental results demonstrate that our approach achieves comparable performance to the current state-of-the-art (SOTA) method, GPV-Pose, under standard conditions. More importantly, in robustness tests with incomplete objects, our method significantly surpasses GPV-Pose. For instance, under a 20% incompleteness ratio, the accuracy of GPV-Pose drops by 61.6% under the \(5^{\circ }2\,\text {cm}\) criterion, whereas our method experiences only a 21.8% reduction.