Unsupervised Selection of Optimal Operating Parameters for Visual Place Recognition Algorithms Using Gaussian Mixture Models

Visual place recognition (VPR) algorithms are a key part of many autonomous systems, but typically consist of many parameters which require non-trivial optimization for a given deployment environment. Being able to automatically select the optimal operating point for parameters within a VPR algorithm would greatly improve the deployability of autonomous systems in real world scenarios. For example, in an aerial context, platform altitude and camera field of view play a critical role in how much of the environment a downward facing camera can perceive. The sensor coverage and its subsequent processing also has significant computational implications. In this letter, we develop an unsupervised system that can predict the performance of a VPR algorithm, using only a limited number of analogous training images. At the core of our approach is the estimation of a recall proxy using Gaussian mixture models and domain-valid assumptions. We develop a robust, intuitive selection criteria to choose the optimal operating point for a deployment environment to show how our system can facilitate automatic parameter selection. Finally, we show how our system can continuously estimate the performance of a VPR system “on-the-fly”. We evaluate our method's effectiveness and generality on both aerial and ground-based real-world datasets. We believe these results will assist in the streamlined deployment of visual localization algorithms in real-world situations.