A Self Attention-Based Feature Learning Framework for Visual Localization

Visual camera localization can be defined as a computer vision problem, that aims to retrieve the pose of a camera given a query image which was originally captured by the camera. This paper proposes a new deep-learning architecture for indoor visual camera localization. Our deep learning framework uses deep spatial points in low-level feature maps as input to precisely model the scene coordinate regression problem. The proposed approach focuses on learning important and invariant visual deep features, which removes the variance between the different level feature representations and combines multi-scale context information. We implement a robust training pipeline that provides an attention-based feature map with straight regulation from the ground truth to enhance pose estimation accuracy. Comprehensive experiments show how our approach achieves comparable performance against previous methods on the indoor benchmark 7-Scenes dataset.