Deep Model-Based 6D Pose Refinement in RGB
This addresses pose refinement for robotics or AR/VR applications, offering a robust, correspondence-free method that handles occlusion and symmetry, though it builds incrementally on contour-based tracking ideas.
The paper tackles the problem of 6D pose refinement in RGB images by proposing a deep neural network that predicts pose updates using a contour-based visual loss, achieving real-time performance and pose accuracies close to 3D ICP without depth data.
We present a novel approach for model-based 6D pose refinement in color data. Building on the established idea of contour-based pose tracking, we teach a deep neural network to predict a translational and rotational update. At the core, we propose a new visual loss that drives the pose update by aligning object contours, thus avoiding the definition of any explicit appearance model. In contrast to previous work our method is correspondence-free, segmentation-free, can handle occlusion and is agnostic to geometrical symmetry as well as visual ambiguities. Additionally, we observe a strong robustness towards rough initialization. The approach can run in real-time and produces pose accuracies that come close to 3D ICP without the need for depth data. Furthermore, our networks are trained from purely synthetic data and will be published together with the refinement code to ensure reproducibility.