Cascaded Parallel Filtering for Memory-Efficient Image-Based Localization

Abstract

Image-based localization (IBL) aims to estimate the 6DOF camera pose for a given query image. The camera pose can be computed from 2D-3D matches between a query image and Structure-from-Motion (SfM) models. Despite recent advances in IBL, it remains difficult to simultaneously resolve the memory consumption and match ambiguity problems of large SfM models. In this work, we propose a cascaded parallel filtering method that leverages the feature, visibility and geometry information to filter wrong matches under binary feature representation. The core idea is that we divide the challenging filtering task into two parallel tasks before deriving an auxiliary camera pose for final filtering. One task focuses on preserving potentially correct matches, while another focuses on obtaining high quality matches to facilitate subsequent more powerful filtering. Moreover, our proposed method improves the localization accuracy by introducing a quality-aware spatial reconfiguration method and a principal focal length enhanced pose estimation method. Experimental results on real-world datasets demonstrate that our method achieves very competitive localization performances in a memoryefficient manner.

Hierarchical Scene Coordinate Classification and Regression for Visual Localization

“…With the established 2D-3D correspondences, a RANSAC [19] optimization scheme is responsible for producing the final pose estimation. The correspondences are typically obtained by matching local features such as SIFT [29], and many matching and filtering techniques have been proposed, which enable efficient and robust city-scale localization [14,25,34,43,52,54]. Image retrieval can also be used for visual localization [1].…”

Learning Camera Localization via Dense Scene Matching

“…Sparse Feature Matching. Methods [8,42,38,48,31,31] based on sparse feature matching build 2D-3D correspon-dences by interest point detection [29,10,12,2,17,32] and local descriptor matching [38,10,40,12,29,6]. Then, poses are estimated by P nP combined with RANSAC.…”

Reference Pose Generation for Long-term Visual Localization via Learned Features and View Synthesis

“…In turn, these 2D-3D matches can be used to estimate the camera pose of the query image by applying an n-point pose solver (Haralick et al 1994;Kukelova et al 2013Kukelova et al , 2010Larsson et al 2017;Albl et al 2016;Kneip et al 2011;Fischler and Bolles 1981) inside a hypothesize-and-verify framework such as RANSAC (Fischler and Bolles 1981) and its variants (Chum and Matas 2008;Lebeda et al 2012;Raguram et al 2013). Research on such 3D structure-based methods has mostly focused on scalability, e.g., by accelerating the 2D-3D matching stage (Li et al 2010(Li et al , 2012Choudhary and Narayanan 2012;Donoser and Schmalstieg 2014;Lim et al 2012;Jones and Soatto 2011;Cheng et al 2019) and the use of image retrieval (Irschara et al 2009;Sattler et al 2012;Sarlin et al 2019;Taira et al 2018;Liu et al 2017;Cao and Snavely 2013), by reducing memory requirements through model compression (Li et al 2010;Cao and Snavely 2014;Camposeco et al 2019;Lynen et al 2015;Dymczyk et al 2015), or by making the pose estimation stage more robust to the ambiguities encountered at scale (Li et al 2012;Zeisl et al 2015;Svärm et al 2017;Toft and Larsson 2016;Alcantarilla et al 2011;Aiger et al 2019).…”

“…This is motivated by our observation that the reference poses for the nighttime images are the least accurate reference poses among the three datasets from Sattler et al (2018). At the same time, the dataset is becoming increasingly popular in the community, e.g., (Sarlin et al 2019;Yang et al 2020;Wang et al 2020;Benbihi et al 2019;Brachmann and Rother 2019;Mishchuk et al 2017;Shi et al 2019;Cheng et al 2019;Revaud et al 2019;Germain et al 2019;Sarlin et al 2020;Zhang et al 2019) have already been evaluated on the dataset. However, our approach is generally applicable and can be applied to other datasets as well.…”