Image Relocalization Based On Deep Feature

Image relocalization is a hot topic in computer vision research,which is the basis and difficulty of AR,SLAM and image-base location recognition.This paper studies the problem of 6-DoF image relocalization based on single monocular RGB image.Unlike the photo location recognition,6-DoF image relocalization requires accurate calculation of image pose.The traditional image relocalization approach is limited by the ability of hand-crafted feature to represent the image,resulting in complicated calculation,poor robustness and low real-time performance.In recent years,deep learning,such as convolution neural network,has made great breakthroughs in image classification,object detection and recognition,and even the Image Caption Generation.After the introduction of the pose estimation based on the 2D feature alignment,this paper analyzes the advantages and disadvantages of the traditional algorithm and the basic principle of the deep convolution neural network,and proposes the deep feature based image pose regression model.A pose regression model using CNN as feature extractor is designed and the problem of lack of pose data set is solved by using pre-train.Since this approach did not model the uncertainty of the problem,we propose a mixture Gaussian idea to improve our approach.After analyzing the similarity between the principle of relocalization task and image classification task,we propose a multi-task learning deep network framework,joint learning image 6 degrees of freedom pose regression and scene classification task.Two different tasks share the same low-level features and train at a same network.The experimental results on the four datasets show that this improved approach greatly reduces the relocalization error of the deep feature based pose regression model.The scene type and probability that our network model output can be used as the confidence of the pose estimation results,so that our approach is useful in practical application.Finally,this paper presents an application that applies pose regression models to city-wide buildings image localization systems.