PanoPR(Spheriverse): A Panoramic Dataset for Pan-Regional Street Scene Understanding

Abstract

Panoramic vision, with its unparalleled 360\textdegree{} environmental perception, provides intelligent systems with a holistic visual understanding of complex scenes. However, due to the lack of diverse and standardized datasets and benchmarks, existing panoramic scene parsing studies remain largely constrained on limited domain, leaving the full potential of panoramic vision for comprehensive understanding underexplored.

To bridge this gap, we present Spheriverse-V1, a diverse panoramic perception dataset tailored to facilitate systematic investigations into panoramic street-view scene understanding. Spheriverse-V1 features: (1) Diverse coverage across urban, suburban, and rural environments with full-day (0–24h) data collection; (2) Multi-task benchmarks including Occupancy Prediction, BEV Perception, and Point Cloud Segmentation; (3) Cross-domain evaluation revealing the challenges and generalization limits of existing perception methods in real-world diverse scenarios; and (4) Comprehensive support through standardized data structures, statistical analyses, and develop tools to facilitate community-driven exploration. Overall, Spheriverse-V1 include xx images and xx sequence, establishes a solid foundation for panoramic street scene parsing and opens new avenues toward unified perception, full-space comprehension, and cross-domain intelligent sensing.
The resources, benchmarks, and toolkits will be made publicly available.