Improved Stacked Hourglass Network for Robust 6D Object Pose Estimation

Abstract

In this article, we introduce an accurate yet robust method to recover the 6D pose of the object from an RGB image. The core of our method is using the farthest point sampling algorithm to design a set of representative keypoints on the object model surface, and then use the improved stacked hourglass network (ISHN) with multi-scale aggregation module to localize them in the 2D image by predicting the keypoints heatmaps. Finally, the PnP algorithm can recover the 6D pose according to the 3D-2D relationship of keypoints. Besides, when the object is partially occluded, we can successfully recover the pose of the object by selecting the most confident keypoints. Our method can simultaneously detect and recover the 6D pose of the instance object in the RGB image without additional post-processing steps. Experimental results show that compared with the state-of-the-art RGB-based pose estimation methods, our method can achieve competitive or more superior performance on two benchmark datasets.