Aerial Lifting: Neural Urban Semantic and Building Instance Lifting from Aerial Imagery

• DOI: 10.48550/arxiv.2403.11812
• Publication Date: 2024-03-18
• Authors: Yuqi Zhang; Guanying Chen; Jiaxing Chen; Shuguang Cui
• Keywords: Aerial imagery; Aerial photos

We present a neural radiance field method for urban-scale semantic and building-level instance segmentation from aerial images by lifting noisy 2D labels to 3D. This is challenging problem due two primary reasons. Firstly, objects in urban exhibit substantial variations size, including buildings, cars, roads, which pose significant challenge accurate segmentation. Secondly, the generated existing methods suffer multi-view inconsistency problem, especially case of images, where each image captures only small portion entire scene. To overcome these limitations, we first introduce scale-adaptive label fusion strategy that enhances varying sizes combining predicted different altitudes, harnessing novel-view synthesis capabilities NeRF. then novel cross-view grouping based on 3D scene representation mitigate labels. Furthermore, exploit reconstructed depth priors improve geometric quality field, resulting enhanced results. Experiments multiple real-world datasets demonstrate our approach outperforms methods, highlighting its effectiveness.