Dynamically Modulating Visual Place Recognition Sequence Length For Minimum Acceptable Performance Scenarios

TL;DR

This work addresses maintaining a target visual place recognition performance under unreliable position sensing by dynamically modulating VPR sequence length. It combines coarse position priors with an MLP regression model trained on appearance-variation features derived from a region around the prior, selecting sequence lengths to exceed the target recall. A Leaky ReLU MSE loss and adjusted mutual information–based feature selection are introduced, and the method is validated across multiple datasets, showing improved consistency over fixed-length approaches while highlighting generalization and the value of non-SOTA, nuanced features. The results underscore practical performance metrics for robust, low-latency localization in real-world robotics, and offer insights into feature selection and transferability across conditions.

Abstract

Mobile robots and autonomous vehicles are often required to function in environments where critical position estimates from sensors such as GPS become uncertain or unreliable. Single image visual place recognition (VPR) provides an alternative for localization but often requires techniques such as sequence matching to improve robustness, which incurs additional computation and latency costs. Even then, the sequence length required to localize at an acceptable performance level varies widely; and simply setting overly long fixed sequence lengths creates unnecessary latency, computational overhead, and can even degrade performance. In these scenarios it is often more desirable to meet or exceed a set target performance at minimal expense. In this paper we present an approach which uses a calibration set of data to fit a model that modulates sequence length for VPR as needed to exceed a target localization performance. We make use of a coarse position prior, which could be provided by any other localization system, and capture the variation in appearance across this region. We use the correlation between appearance variation and sequence length to curate VPR features and fit a multilayer perceptron (MLP) for selecting the optimal length. We demonstrate that this method is effective at modulating sequence length to maximize the number of sections in a dataset which meet or exceed a target performance whilst minimizing the median length used. We show applicability across several datasets and reveal key phenomena like generalization capabilities, the benefits of curating features and the utility of non-state-of-the-art feature extractors with nuanced properties.

Figures (7)

• Figure 1: Overview of our method. Our approach modulates VPR sequence length to exceed a target localization performance using coarse position priors and an MLP regression model. The multi-coloured bar plot shows how the sequence length changes over the dataset (see colour scale) to exceed the target localization performance (black dotted line). The green and blue bar plots show the improvements made over fixed sequence lengths across several datasets by increasing the proportion of the dataset that exceeds the target recall or reducing the median sequence length for the same performance (Explained in Fig. \ref{['fig:oracomps']} and Sect. \ref{['subsec:resultsmain']}).
• Figure 2: a) Each line represents the performance of a sequence length in a particular chunk of the dataset. The dotted line represents the chosen target recall value for this dataset. b) The smallest sequence length for each chunk required to exceed the target recall; or the sequence length closest to the target performance in the case it is not exceeded.
• Figure 3: a) The full feature vector $v_{n}$ is reduced to select features, $v_{p}$, using a threshold for the correlation with target sequence length, $s$. b) The Leaky ReLU activation used to control the penalty for over and under predictions separately.
• Figure 4: Example images from the datasets used in our work. Top Row: RobotCar Dusk, Overcast and Sunny. 2nd Row: Nordland Fall, Spring and Summer. 3rd and 4th Rows: query and reference images respectively from the Ford 1 2017, Ford 3, Ford 4 and Ford 1 2022 datasets.
• Figure 5: Delta improvements of our approach over fixed sequence lengths when matching its median sequence length (Top) or consistency (Bottom) metrics. Dataset indices follow order in Table \ref{['tab:main']}.