Camera Re-localization by Training Multi-dataset Simultaneously via Convolutional Neural Network

With the advances of Convolutional Neural Networks (CNN) in computer vision, rapid progresses have been taken in camera re-localization. In this paper we propose a new network, a multi-dataset simultaneously training network (MdNet) to relocate camera pose from an RGB image. Moreover, we propose to construct new loss functions to learn camera pose, image segmentation and images depth maps from the multi-datasets. Compared with PoseNet Geometric Loss in street dataset, position and orientation accuracy are increased by 52% and 35% respectively. Experiment shows that our method outperforms other prior similar works in large-scale scenarios and is more robust under different season or time conditions.