KappaPlace: Uncertainty-Aware Visual Place Recognition via Prototype-Anchored Supervision
Researchers have unveiled KappaPlace, a new framework aimed at enhancing visual place recognition (VPR) by incorporating uncertainty awareness. It employs a Prototype-Anchored supervision method that uses latent class representatives as targets for a probabilistic goal. By viewing image descriptors as von Mises-Fisher (vMF) variables, a simplified module calculates the concentration parameter to capture aleatoric uncertainty. Furthermore, a fresh match-level formulation evaluates the reliability of specific query-reference pairs, addressing the limitations of existing query-focused techniques. The performance of this innovative approach has been evaluated on five different benchmarks.
Key facts
- KappaPlace is a framework for uncertainty-aware visual place recognition.
- It uses Prototype-Anchored supervision with latent class representatives.
- Image descriptors are modeled as von Mises-Fisher (vMF) variables.
- A lightweight module predicts concentration parameter for aleatoric uncertainty.
- A novel match-level formulation quantifies query-reference pair reliability.
- Existing VPR uncertainty methods are typically query-centric.
- Evaluated across five diverse benchmarks.
- Published on arXiv as paper 2605.19435.
Entities
Institutions
- arXiv