Microscale Precision of 6DOF Localization Rectification of Low-End Stereo Vision Using Deep Learning

Abstract

This paper proposes the deployment of a deep feedforward neural network (DFF) for the estimation of an object's 6D pose with respect to a robot arm. DFF is also deployed as a data rectification stage. The proposed technique is able to achieve microscale precision using customized low-end stereo vision system, given the 3D location of each of the object's four bounding corners. The 4 bounding corners' 3D locations are estimated from the subpixel localization accuracy of each. The paper also proposes a building block for convolutional neural networks for object detection, which is based on utilization of skipping connections that concatenates early feature maps to later layers, instead of adding them like in Residual network (ResNet). The proposed building block shows better results than the building block of ResNet in terms of object detection accuracy. For minimizing the use of memory resources, this paper is recommending the use of batch normalization layer at the beginning of convolutional neural network architectures, empirically, it dispenses the need for data normalization.