"Cross-View 3D Object detection"

"We introduce a large-scale cross-view benchmark that establishes precise geometric alignment between ground-level street-view imagery and overhead satellite observations for urban 3D perception. The dataset is constructed upon widely used autonomous-driving datasets and augments each street-view frame with a geographically aligned satellite image crop centered at the camera location and orientation. Such cross-view pairing enables consistent correspondence between object-level semantics in the ground view and static scene structures observable from the aerial perspective.The benchmark supports multiple 3D perception tasks under severe viewpoint discrepancies, including monocular 3D object detection and spatial localization, and is designed to facilitate research on cross-view representation learning and geometric reasoning. All samples are provided with unified coordinate systems, standardized preprocessing pipelines, and evaluation protocols to ensure reproducibility and fair comparison. By explicitly leveraging complementary information from satellite imagery, this dataset serves as a challenging testbed for advancing robust 3D perception in complex urban environments."