The 360Loc dataset contains 4 diverse indoor and outdoor scenes featuring symmetrical, repetitive structures and moving objects. It was collected using a portable 360-camera-lidar platform, and the ground truth 6DoF poses were generated through a series of optimizations involving lidar mapping, bundle adjustment, and point cloud registration.
To enable cross-device visual localization, the dataset includes not only 360° reference images, but also query frames from pinhole, fisheye, and 360° cameras. A virtual camera approach was introduced to generate high-quality lower-FoV images from the 360° views, ensuring a fair comparison of performance among different query types.
The authors extend feature-matching-based and absolute pose regression pipelines to support omnidirectional visual localization. The virtual camera method is used to reduce the domain gap between query and reference images, improving the performance of image retrieval and absolute pose regression. Extensive evaluations demonstrate the advantages of 360° cameras in reducing ambiguity in visual localization on scenes with symmetric or repetitive features, as well as the effectiveness of the virtual camera approach in enhancing cross-device visual localization.
To Another Language
from source content
arxiv.org
Thông tin chi tiết chính được chắt lọc từ
by Huajian Huan... lúc arxiv.org 04-09-2024
https://arxiv.org/pdf/2311.17389.pdfYêu cầu sâu hơn