6D Pose Estimation Based on the Adaptive Weight of RGB-D Feature

摘要

In the task of 6D pose estimation by RGB-D image, the crucial problem is how to make the most of two types of features respectively from RGB and depth input. As far as we know, prior approaches treat those two sources equally, which may overlook that the different combinations of those two properties could have varying degrees of impact. Therefore, we propose a Feature Selecting Mechanism (FSM) in this paper to find the most suitable ratio of feature dimension from RGB image and point cloud (converted from depth image) to predict the 6D pose more effectively. We first conduct artificial selection in our Feature Selecting Mechanism (FSM) to prove the potential for the weight of the RGB-D feature. Afterward, the neural network is deployed in our FSM to adaptively pick out features from RGB-D input. Through our experiments on the LINEMOD dataset, YCB-Video dataset, and our multi-pose synthetic image dataset, we show that there is an up to 2% improvement in the accuracy by utilizing our FSM, compared to the state-of-the-art method.