Real-Time Grasp Detection Using Efficient Channel Attention

To realize the intelligence of robots, robots need to have the ability to grasp unknown objects. In this paper, a novel grasp detection network is proposed, named Efficient Channel Attention Grasp Network (ECA-GraspNet). ECA-GraspNet uses an encoder-decoder architecture to fuse feature information from different layers using a Feature Pyramid Network (FPN). The network is able to actively focus on convolutional layers that are more useful for grasp detection by embedding Efficient Channel Attention (ECA). The proposed network is able to generate pixel-by-pixel grasp poses from RGB-D images at real-time speed. We evaluate the proposed network on the public Cornell dataset and achieve an accuracy of 97.7%.