Text-Pose Estimation in 3D Using Edge-Direction Distributions

This paper presents a method for estimating the orientation of planar text surfaces using the edge-direction distribution (EDD) extracted from the image as input to a neural network. We consider canonical rotations and we developed a mathematical model to analyze how the EDD changes with the rotation angle under orthographic projection. In order to improve performance and solve quadrant ambiguities, we adopt an active-vision approach by considering a pair of images (instead of only one) with a slight rotation difference between them. We then use the difference between the two EDDs as input to the network. Starting with camera-captured front-parallel images with text, we apply single-axis synthetic rotations to verify the validity of the EDD transform model and to train and test the network. The presented text-pose estimation method is intended to provide navigation guidance to a mobile robot capable of reading the textual content encountered in its environment.