Drone-View Building Identification by Cross-View Visual Learning and Relative Spatial Estimation

Chun-Wei Chen, Yin-Hsi Kuo, Tang Lee, Cheng-Han Lee, Winston H. Hsu

Abstract: Drones become popular recently and equip more sensors than traditional cameras, which bring emerging applications and research. To enable drone-based applications, providing related information (e.g., building) to understand the environment around the drone is essential. We frame this drone-view building identification as building retrieval problem: given a building (multimodal query) with its images, geolocation and drone's current location, we aim to retrieve the most likely proposal (building candidate) on a drone-view image. Despite few annotated drone-view images to date, there are many images of other views from the Web, like ground-level, street-view and aerial images. Thus, we propose a cross-view triplet neural network to learn visual similarity between drone-view and other views, further design relative spatial estimation of each proposal and the drone, and collect new drone-view datasets for the task. Our method outperforms triplet neural network by 0.12 mAP. (i.e., 22.9 to 35.0, +53% in a sub-dataset [LA])