Object Pose Estimation Based on Improved YOLOX Algorithm

Abstract

This paper proposes a deep learning method for object recognition and pose estimation only through a single RGB image. We extend the YOLOX algorithm, which has excellent performance in the field of 2D object detection, making it suitable for 6DoF pose estimation scenarios in a natural way. The model is designed to predict the position of key points in the image, which are used to establish the correspondence between the object model and the scene image. After that, the object pose can be obtained by PnP algorithm. In order to obtain labeled data for model training effectively, this paper designs a data augmentation method combining offline expansion and online augmentation, which can improve the generalization ability of the model. The results of experiments on the LINEMOD dataset demonstrate the effectiveness of our method, which is significantly competitive with other methods using similar implementation ideas.