AutoCompass: Accurate Visual Localization on Public Maps by Learning from Weak Labels

AutoCompass is a supervision method that trains neural map matchers for 3‑DoF visual localization using weak, noisy absolute pose labels. The approach demonstrates that heading labels can be omitted—models learn accurate headings from raw GPS alone—and that defining a tolerance region around GPS improves positional accuracy. When relative pose information from SLAM or SfM is available, it further enhances training, leading to consistent performance gains over models trained with conventional absolute pose labels.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv Computer Vision.

• benchmarks