Predicting world coordinates of pixels in RGB images using Convolutional Neural Network for camera relocalization

Convolutional Neural Networks (CNNs) have achieved great successes in many computer vision tasks and have been applied to pose regression for camera relocalization. Traditional Simultaneously Localization and Mapping (SLAM) approaches use correspondences between camera coordinates and world coordinates to estimate camera pose. In this paper, we present a new camera relocalization method including pixels' world coordinates regression with CNNs and camera pose optimization. We also explore the different characteristics of CNNs and SCoRe Forests on world coordinates regression. Experiments show that our approach has larger camera relocalization error but better performance on predicting world coordinates of pixels compared to SCoRe Forests.