Understanding the Role of Visual Place Recognition for 3D Reconstruction

Structure from Motion (SfM) is the task of building a 3D reconstruction of the surrounding space, extracting information from 2D images. Visual Place Recognition (VPR), on the other hand, is the task of recognising and matching images depicting the same place, possibly from different points of view. A crucial step in SfM pipelines, allowing both to increase efficiency and boost performance, is matching similar images with VPR algorithms before the reconstruction is initialised. This thesis aims at investigating the role of VPR in the process of building a 3D reconstruction, pointing out the features that a place recognition method should have to work fine in this task. This is achieved carrying out numerical experiments over images, sourced from the Internet, of three large scale scenes. An original evaluation strategy, based on point reprojection on a set of test images, is proposed and used to compare performances of various VPR methods and strategies in a Structure from Motion pipeline.