FidelityBEV: A Fidelity-Preserving Network for Monocular Bird's-Eye-View Semantic Segmentation

Monocular Bird's-Eye View (BEV) semantic segmentation is a critical task in autonomous driving perception[1], with its performance bottleneck residing in the view transformation pipeline. This paper systematically demonstrates that prevailing methods suffer from a cascading fidelity loss, where information is distorted at the semantic, geometric, and structural levels during the transformation process. To address this issue, we propose FidelityBEV, a novel network designed for end-to-end fidelity preservation. This network rectifies the information flow through three synergistic modules: (1) a Semantic-Structural Synergy Module (S3 Module) to enhance the semantic fidelity of source information; (2) an Uncertainty Gating Unit (UGU) to preserve geometric fidelity under uncertainty; and (3) a Vertical Context Aggregator (VCA) to ensure structural fidelity during the projection process. On the KITTI-360 benchmark, FidelityBEV achieves 41.66% in mean Intersection over Union (mIoU), marking a substantial improvement of 6.43 percentage points over the baseline.