Unified Retrieval and Reranking Paradigm for Aerial-Ground Cross-Source 3D Place Recognition

3D place recognition aims to localize a query point cloud by matching it to the reference point cloud database. Previous approaches have mainly been developed for single-source place recognition, but significantly degraded to cross-source place recognition. This occurs because cross-source point clouds inherently exhibit point densities and perspective differences. In this paper, we propose an aerial-ground cross-source 3D place recognition method that handles retrieval and reranking simultaneously. In the retrieval stage, we employ a CRFormer with circle and radical partitions to enlarge respective fields and efficiently enhance the information communication of long-range context. In the reranking stage, we further introduce a correlation-guided selection mechanism to discard redundant local features with consideration of the perspective differences. Experimental results demonstrate that the proposed method attains state-of-the-art performance on the CS-Campus3D benchmark, surpassing previous approaches by a significant margin of 6.8% and 6.98% in AR@1% and AR@1, respectively.