LaLaLoc: Latent Layout Localisation in Dynamic, Unvisited Environments

TL;DR

LaLaLoc is a novel localization method that uses latent representations of room layouts to accurately determine position without prior scene visitation, even with significant scene changes.

Contribution

LaLaLoc introduces a new approach leveraging shared embedding spaces and cross-modal pose optimization for robust, prior-free localization in dynamic environments.

Findings

Achieves 8.3cm localization accuracy in domestic settings.

Robust to scene changes like furniture rearrangement.

Operates without prior scene visitation.

Abstract

We present LaLaLoc to localise in environments without the need for prior visitation, and in a manner that is robust to large changes in scene appearance, such as a full rearrangement of furniture. Specifically, LaLaLoc performs localisation through latent representations of room layout. LaLaLoc learns a rich embedding space shared between RGB panoramas and layouts inferred from a known floor plan that encodes the structural similarity between locations. Further, LaLaLoc introduces direct, cross-modal pose optimisation in its latent space. Thus, LaLaLoc enables fine-grained pose estimation in a scene without the need for prior visitation, as well as being robust to dynamics, such as a change in furniture configuration. We show that in a domestic environment LaLaLoc is able to accurately localise a single RGB panorama image to within 8.3cm, given only a floor plan as a prior.