Revisiting Semantic Class Uncertainties for Robust Visual Place Recognition

Published: 09 Apr 2024, Last Modified: 10 Apr 2024ICRA 2024: Back to the FutureEveryoneRevisionsBibTeXCC BY 4.0
Keywords: visual place recognition, foundation model, semantic segmentation, uncertainty
TL;DR: Semantic Classes in VPR is not optimally handled by assigning weights, and its rather a conditional probability distribution.
Abstract: Visual place recognition has been challenging and crucial in real-world applications such as autonomous navigation and vision-based robot missions. The introduction of foundation models has greatly enhanced the accuracy of vision-based algorithms and visual place recognition, expanding to the methods that utilize semantic information from images. In this workshop paper, we revisit the features of semantic classification in the visual place recognition process to discuss how to deal with outcomes of semantic segmentation during visual place recognition. By showing that the semantic labels are not uniformly distributed, we propose to handle the uncertainty of semantic classes as a bivariate distribution that depends on the class and the assigned localization clusters instead of commonly used class-level confidences. Utilizing a powerful foundation model capable of language-image similarity evaluation, we evaluate and show the distributions of semantic class activations in the public datasets.
Submission Number: 9