360Loc: A Dataset and Benchmark for Omnidirectional Visual Localization with Cross-device Queries

Portable 360^∘ cameras are becoming a cheap and efficient tool toestablish large visual databases. By capturing omnidirectional views of ascene, these cameras could expedite building environment models that areessential for visual localization. However, such an advantage is oftenoverlooked due to the lack of valuable datasets. This paper introduces a newbenchmark dataset, 360Loc, composed of 360^∘ images with ground truthposes for visual localization. We present a practical implementation of360^∘ mapping combining 360^∘ images with lidar data to generatethe ground truth 6DoF poses. 360Loc is the first dataset and benchmark thatexplores the challenge of cross-device visual positioning, involving360^∘ reference frames, and query frames from pinhole, ultra-wide FoVfisheye, and 360^∘ cameras. We propose a virtual camera approach togenerate lower-FoV query frames from 360^∘ images, which ensures a faircomparison of performance among different query types in visual localizationtasks. We also extend this virtual camera approach to feature matching-basedand pose regression-based methods to alleviate the performance loss caused bythe cross-device domain gap, and evaluate its effectiveness againststate-of-the-art baselines. We demonstrate that omnidirectional visuallocalization is more robust in challenging large-scale scenes with symmetriesand repetitive structures. These results provide new insights into 360-cameramapping and omnidirectional visual localization with cross-device queries.