Rethinking map densification for Visual Place Recognition

Great work by Mubariz Zaffar, PhD in the 3D Urban Understand (3DUU) DAI Lab, which resulted in an article publication in the top robotics journal Transactions on Robotics!

Visual Place Recognition aims to determine the location of a query image by comparing its image descriptor to those of a set of reference images with known geographic locations. However, the localization accuracy is limited by how densely these reference images are distributed in the physical space. Extending the reference set to include more intermediate locations can improve the localization accuracy, but simply collecting more images at the missing locations is not always possible. Recent work has attempted to construct a 3D model of the scene from the available reference images to generate novel views for the missing locations, but accurate 3D reconstruction from only one or a few images remains challenging and computationally expensive. In this article, we show that it is instead possible to directly regress the descriptors for the missing locations from the available reference descriptors of nearby locations, without the need of synthesizing the actual images at those poses. In our experiments, we observe ~30% increase in localization accuracy by densifying standard benchmarks, and show that it can complement accuracy improvements by the popular coarse-to-fine localization.

"CoPR: Towards Accurate Visual Localization With Continuous Place-descriptor Regression", M. Zaffar, L. Nan, J.F.P. Kooij, Transactions on Robotics (T-RO), 2023, doi: 10.1109/TRO.2023.3262106