Attention-Driven Object Encoding and Multiscale Contextual Perception for Improved Cross-View Object Geo-Localization

Abstract

Cross-view object geo-localization (CVOGL) is essential for applications like navigation and intelligent city management. By identifying objects in street-view/drone-view and precisely locating them in satellite imagery, more accurate geo-localization can be achieved compared to retrieval-based methods. However, existing approaches fail to account for query object shape/size and significant scale variations in remote sensing images. To address these limitations, we propose an attention-driven multiscale perception network (AMPNet) for cross-view geo-localization. AMPNet employs an attention-driven object encoding (ADOE) based on segmentation, which provides prior information to enable learning more discriminative representations of the query object. In addition, AMPNet introduces a cross-view multiscale perception (CVMSP) module that captures multiscale contextual information using varying convolution kernels, and applies an MLP to enhance channel-wise feature interactions. Experimental results demonstrate that AMPNet outperforms state-of-the-art methods in both ground-to-satellite and drone-to-satellite object localization tasks on a challenging dataset.