Hybrid Representation Guided Geometric Regression Network

Estimating 6D pose from a single RGB image is a fundamental task in computer vision. Traditional approaches typically involve solving for the 6D pose by first establishing dense 2D-3D correspondences and then applying a PnP algorithm. While these methods achieve high accuracy, they suffer from two significant drawbacks. First, they often rely on proxy objectives, such as correspondence regression, for training, which may not accurately reflect the optimized 6D pose error. Second, these methods are non-differentiable with respect to the estimated 6D pose, which limits the learning process. In this paper, we propose a novel deep learning method for 3D object detection and 6D pose estimation from RGB images. Our approach, named the Hybrid Representation Guided Geometric Regression Network, leverages a hybrid representation that allows pose regression to exploit richer and more diverse features. A simple network predicts keypoints, edge vectors, and symmetry correspondences-three different intermediate geometric cues that provide ample geometric constraints for pose regression. By utilizing geometric guidance based on hybrid intermediate representations, our method enhances the performance of direct 6D pose regression. This approach achieves direct pose regression, and the incorporation of diverse geometric information improves the accuracy of pose estimation.