YOLO-6D+: Single Shot 6D Pose Estimation Using Privileged Silhouette Information
Jia Kang, Wenjun Liu, Wenzhe Tu, Lu Yang
Abstract
The task of estimating the 6D pose of the object from a single RGB image is important for augmented reality and robotic grasping applications. In this work, we introduce YOLO-6D+, a new end-to-end deep network for 6D object pose estimation. In particular, we propose a novel silhouette prediction branch that outputs the predicted segmentation mask in our network, which can force underlying features to learn the silhouette information of the object. Furthermore, we introduce edge restrain loss, a new loss function that focuses on constraining the 3D shape of an object. We use a two-stage method: we predict 2D keypoints firstly and then 6D pose is estimated using the PnP algorithm. On the public LINEMOD dataset, we demonstrate the proposed approach can outperform the state-of-the-art YOLO-based single shot pose estimation approach [1] by 4.09% and 11.72% under the 2D projection metric and the ADD(-s) metric respectively.