OneBEV++: Towards Unifying Bird's-Eye-View Semantic Mapping with Panoramas

Area of Science:

• Computer Vision
• Robotics
• Autonomous Systems

Background:

• Bird's-Eye-View (BEV) perception is vital for scene understanding but limited by sensor constraints.
• Narrow field-of-view sensors (indoor) and multi-camera setups (outdoor) present significant challenges.
• Existing methods struggle with omni-range perception from a single image.

Purpose of the Study:

• To develop a novel pipeline for panoramic-to-BEV semantic mapping.
• To enable omni-range perception using a single panoramic image.
• To create benchmarks for evaluating panoramic BEV in diverse settings.

Main Methods:

• Proposed a new framework, OneBEV++, for generating BEV semantic maps from single panoramic images.
• Extended existing datasets to create four benchmarks for indoor and outdoor panoramic BEV.
• Evaluated OneBEV++ against prior methods like 360BEV and OneBEV.

Main Results:

• OneBEV++ demonstrated superior performance across four new panoramic BEV datasets.
• Consistent accuracy gains and improved efficiency were observed compared to existing methods.
• The framework successfully generated BEV semantic maps from single panoramic images in both indoor and outdoor environments.

Conclusions:

• OneBEV++ offers a simplified yet high-performance solution for omni-range perception.
• The method enables scalable applications for autonomous vehicles and indoor robots.
• This research paves the way for more integrated cross-environment autonomous systems.