Estimating Grasping Patterns from Images Using Finetuned Convolutional Neural Networks

Identification of suitable grasping pattern for numerous objects is a challenging computer vision task. It plays a vital role in robotics where a robotic hand is used to grasp different objects. Most of the work done in the area is based on 3D robotic grippers. An ample amount of work could also be found on humanoid robotic hands. However, there is negligible work on estimating grasping patterns from 2D images of various objects. In this paper, we propose a novel method to learn grasping patterns from images and data recorded from a dataglove, provided by the TUB Dataset. Our network retrains, a pre-trained deep Convolutional Neural Network (CNN) known as AlexNet, to learn deep features from images that correspond to human grasps. The results show that there are some interesting grasping patterns which are learned. In addition, we use two methods, Support Vector Machines (SVM) and hotelling’s T2 test to demonstrate that the dataset does include distinctive grasps for different objects. The results show promising grasping patterns that resembles actual human grasps