Integrating Visual and Geometric Consistency for Pose Estimation

In this work, we tackle the problem of estimating the relative pose between two cameras in urban environments in the presence of additional information provided by low quality localization and orientation sensors. An M-estimator based approach provides an elegant solution for the fusion between inertial and vision data, but it is sensitive to the prior importance of the visual matches between the two views. In addition to using cues extracted from local visual similarity, we propose to rely at the same time on learned associations provided by the global geometrical coherence. A conservative weighting scheme for combining the two types of cues has been proposed and validated successfully on an urban dataset.