On the Impact of Dataset Quality in Multiview BEV Mapping

Abstract

Bird s eye view (BEV) mapping from multicamera systems has emerged as an important form of ground-level remote sensing for constructing geospatial representations in autonomous navigation and urban analytics. Although recent advances in deep learning architectures have expanded the capabilities of BEV mapping, the impact of data-quality factors on mapping performance remains insufficiently explored. In this work, we investigate how fundamental dataset characteristics, such as spatial coverage, ground sampling resolution, annotation quality, and cross-domain differences impact BEV map accuracy, boundary fidelity, and robustness. Using the nuScenes dataset as a primary case study, we systematically vary these data-quality attributes and quantify their effects using geospatial accuracy metrics. To ensure the generality of our findings, all experiments are conducted using two architecturally distinct models, BEVCar and SimpleBEV, and the observed trends are consistent across both. Our results show that reducing BEV map coverage consistently improves accuracy by limiting predictions to visible regions. In addition, we found that resolution and class granularity jointly affect boundary quality; finer annotations at higher resolutions enhance precision for small structures. Once the previous quality factors were evaluated, we created a synthetic dataset under conditions similar to nuScenes to assess generalization under domain shifts and evaluate cross-domain performance between real and synthetic environments. We highlighted that such data augmentation reduces the tendency to hallucinate unseen objects. By quantifying the data-quality drivers of mapping performance, this work offers general insights for the design, curation, and quality assessment of AI-based remote sensing datasets used for geospatial mapping and perception, which are essential for sustainable mobility and resilient urban planning.