Estimate 6D Object Pose for Robotic Grasping Based on Efficient Channel Attention Mechanism

Robots with intelligent grasping can replace humans to complete repetitive tasks. In real scenes with diverse objects randomly placed, it is difficult for robotic grasping due to wrong localization of objects. To address this problem, we propose a pose estimation method for robotic grasping that can solve complex situations such as occlusion. Estimating the 6D pose of objects from an RGB image is a fundamental and challenging task in computer vision. Our method first establishes dense 2D-3D correspondences between 2D image plane and 3D object model, and then regresses the 6D object pose by a neural network. The experimental results show that our proposed method can estimate accurate poses of multiple objects and outperform existing similar methods on three different benchmark datasets.