2017/07/31 by Zakaria Laskar, Laskar, Zakaria, Iaroslav Melekhov +5 · 6 citations
Engineering · Computer Science · Earth and Planetary Sciences · #Robotics and Sensor-Based Localization #Advanced Vision and Imaging #3D Surveying and Cultural Heritage
paper · pdf · doi:10.48550/arxiv.1707.09733
We propose a new deep learning based approach for camera relocalization. Our\napproach localizes a given query image by using a convolutional neural network\n(CNN) for first retrieving similar database images and then predicting the\nrelative pose between the query and the database images, whose poses are known.\nThe camera location for the query image is obtained via triangulation from two\nrelative translation estimates using a RANSAC based approach. Each relative\npose estimate provides a hypothesis for the camera orientation and they are\nfused in a second RANSAC scheme. The neural network is trained for relative\npose estimation in an end-to-end manner using training image pairs. In contrast\nto previous work, our approach does not require scene-specific training of the\nnetwork, which improves scalability, and it can also be applied to scenes which\nare not available during the training of the network. As another main\ncontribution, we release a challenging indoor localisation dataset covering 5\ndifferent scenes registered to a common coordinate frame. We evaluate our\napproach using both our own dataset and the standard 7 Scenes benchmark. The\nresults show that the proposed approach generalizes well to previously unseen\nscenes and compares favourably to other recent CNN-based methods.\n