Original title: AutoCompass: Accurate Visual Localization on Public Maps by Learning from Weak Labels
Javier Tirado-Garín, Alan Savio Paul, Shuai Chen, Axel Barroso-Laguna +4 · September 2026
Cameras can help find where you are on a 2D map. To do this, models learn from large sets of photos paired with GPS locations and headings. The problem is that these location and direction labels are often noisy and full of errors, which hurts the system. The researchers built AutoCompass to train models using these imperfect labels. It works in three ways. First, it drops heading labels entirely; models trained on raw GPS learn correct headings on their own. Second, it sets an error tolerance zone around noisy GPS coordinates to boost position accuracy. Third, it adds relative movement data between photos. Across driving and body-worn camera tests, AutoCompass beat standard models trained on strict absolute labels.
Imagine you are hiking with a paper map, but your GPS is blurry about your exact spot, and no one tells you which way you are facing. You could keep stopping to ask for directions, which slows you down. Here, AutoCompass lets the map system learn to guess your facing from the blurry GPS alone and uses clues from nearby landmarks to pinpoint your location more accurately.
Self-driving cars and AR apps need precise location and direction to work safely. AutoCompass lets them learn from cheap, noisy GPS data instead of requiring expensive, perfect labels.
The abstract does not specify limitations, but it focuses on driving and egocentric benchmarks, so performance on other scenarios is not covered. It also assumes the availability of relative poses via SLAM or SfM for part of the approach.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
AutoCompass: Accurate Visual Localization on Public Maps by Learning from Weak Labels — Javier Tirado-Garín, Alan Savio Paul, Shuai Chen et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.