Learning-based estimation of 6-DoF camera poses from partial observation of large objects for mobile AR

Abstract

We propose a method that estimates 6-DoF camera pose from a partially visible large object, by exploiting information of its subparts that are detected using a state-of-the-art convolutional neural network (CNN). The trained CNN outputs two-dimensional bounding boxes around subparts and associated classes. Information from detection is then fed to a deep neural network that regresses to camera's 6-DoF poses. Experimental results show that the proposed method is more robust to occlusions than conventional learning-based methods.