For navigating the complex surrounding, we envision a proactive and adaptive drone-based building identification technology — recognizing the key buildings in arbitrary deployed cities. However, navigating in the complex city will be a huge problem even with the state-of-the-art sensing and mapping technologies as flying in the sky is hugely different from driving on the streets.

新的嘗試，有沒有可能無人機出廠之後，可以在任一個都市飛行過程中，動態、主動地辨識出主要的地標或是建築物？

We frame this drone-view building identification as “few-shot learning problem” for the major buildings in the city. The identification process is on the fly and hopes to adapt to different cities whenever deployed.

As the drone navigates the city, the (surrounding) buildings of interest are retrieved from the Web (or maps) provided current drone location. We associate the building proposals (in the drone view) with the candidates (in the Web or maps) by the (learned) visual and sensor similarities. In general, the buildings are with few user-contributed or even street-view photos. However, the view is completely different from the drone view. We designed a cross-view learning neural networks to address the problem. We also investigate the impacts from other sensor information (e.g., GPS, compass, etc.)

To our best knowledge, this is probably the first work to address the problem. There should other promising methods for further improving the newly formed problem. We also make the drone-view building identification dataset: “DroneViewBI” public for the academic usage. Feel free to comment and enjoy the carefully collected dataset. The work will also be presented in the CVPR 2018 workshop.