Скачать или смотреть SparseLoc: Localization in Sparse Topometric Maps

We propose city-scale localization using sparse topometric maps augmented with foundational features. We achieve sparsity by registering language-aligned landmark cluster centroids exploiting vision-language models. During the query run, we perform 6 DoF pose estimation through geometric and semantic matching following the Monte Carlo localization scheme. We further propose a novel pose refinement step that follows pose graph optimisation principles. Through the experiments, we show that our framework achieves kilometre-scale localization with navigation-ready accuracy while using less than 1% of the dense map built by standard SLAM techniques.

Project Page: https://reachpranjal.github.io/sparse...