A Novel 6D Pose Estimation Method for Indoor Objects Based on Monocular Regression Depth

Estimating the 6D pose of known objects has attracted a lot of research attention since it is important for intelligent robot manipulation and virtual reality. In this paper, in order to improve the performance of 6D pose estimation using RGB image, we propose a novel 6D object pose estimation framework based on monocular regression depth. To get the depth map of the RGB image, we use a U-Net framework to regress the depth information effectively. For 6D pose estimation, our approach has two steps, which first uses the Convolutional Neural Network (CNN) to extract the feature of the RGB data and the regressed depth data, respectively. Then we adopt a pixel-wise dense fusion module to extract the dense feature to estimate the final 6D object pose. The experiments conducted on the YCB-Video dataset have demonstrated the effectiveness of our approach, and the experiment results show that the proposed method achieves an outstanding performance using only RGB images.