Efficient-NVR: accurate object 6D pose estimation via enhancing semantic representation

Estimating the 6-degree-of-freedom (6Dof) pose for objects is a fundamental task in vision-based measurement. It offers targets' 3D position and orientation information with respect to the camera, which is valuable in various applications, such as robotics, autonomous driving, and augmented reality. Among different approaches, monocular vision methods have the advantage of being flexible and economical. It extracts features from a single RGB image and matches them with the corresponding parts of the target's known 3D model. Recently, regression methods directly predict objects' 6Dof pose have dominated this field by leveraging Convolution Neural Networks (CNN) and learning from tremendous data to extract semantic features. The previous method that leverages objects' surface normal vectors to disentangle rotation estimation from translation achieves superior performance. However, it adopts a backbone network to extract orientation and position features from the input image simultaneously. Therefore, the backbone network restricts the method's overall performance. In this paper, we illustrate this problem and adopt an advanced backbone network as well as a Feature Pyramid Network (FPN) to enhance the feature-extracting capability of our method. We conduct various experiments and ablation studies to demonstrate the outperformance and effectiveness of our newly proposed network, namely Efficient-NVR. Notably, it surpasses state-of-the-art methods on the Linemod benchmark by obtaining 1.3% more accuracy than the baseline.