Multiview Aerial Visual Recognition (MAVREC): Can Multi-View Improve Aerial Visual Perception?

Figure 1. Illustration of the geography-aware model using our proposed MAVREC dataset (green box) collected in the rural and urban European landscape vs. the conventional aerial object detector (blue box) pretrained only on aerial images from VisDrone [91] captured in Asia. The conventional approach fails to detect aerial objects from the MAVREC dataset precisely. In contrast, our object detector pretrained on the ground and aerial images from the MAVREC dataset contextualizes the object proposals of that specific geography and enhances the aerial visual perception, thus outperforming other object detectors pre-trained on popular ground-view dataset (MS-COCO [44]) or other aerial datasets collected from different geographies; also, see Figure 5 .