CycleBEV: Regularizing View Transformation Networks via View Cycle Consistency for Bird's-Eye-View Semantic Segmentation

Researchers propose CycleBEV, a new regularization framework that improves bird's-eye-view semantic segmentation for autonomous driving by using cycle consistency to enhance view transformation networks. The method shows significant improvements up to 4.86 mIoU without increasing inference complexity.

• →CycleBEV introduces an inverse view transformation network that maps BEV segmentation back to perspective view for training regularization.
• →The framework achieves improvements of up to 0.74, 4.86, and 3.74 mIoU for drivable area, vehicle, and pedestrian classes respectively.
• →The method works with existing view transformation models across three major paradigms without adding inference overhead.
• →Cycle consistency principles from image distribution modeling are successfully applied to autonomous driving perception tasks.
• →Implementation code is publicly available and the method was validated on the large-scale nuScenes dataset.