DroneViewBI

by Chun-Wei Chen,Yin-Hsi Kuo,Tang Lee,Cheng-Han Lee,Winston HsuUnknown

DroneViewBI

** Drone-BR ** We collect drone-view images with our drone (DJI 3) by designing and using an app for automatically recording all drone sensors including GPS tracker, compass and altimeter while capturing images. We fix the angle of depression as 20°, raise the drone to 40m to 90m height and capture images with multiple buildings. Each image consists of 1-12 buildings as queries which have corresponding bounding boxes on drone-view images. We collect 80 images and annotate 108 buildings with bounding boxes on them at 3 locations under different weather. Besides, with Google Places API, we gather other information for these buildings including name, latitude and longitude. We gather ground-level, street-view and aerial images for each building. Since the same buildings may appear in different images, there are 585 building queries in total, which is Drone-BR. Since buildings may be occluded or in low resolution, it requires great effort to annotate them with their corresponding high-resolution drone-view images (mostly 3840x2160). In our setting, we search each building query in 200 proposals on a drone-view image, which means the Drone-BR can be also described as 585 queries with 16,000 (200*80) sub-images for searching. Our experiment results show that our method can perform better with limited training data and it is scalable because the building identification for each drone-view image is performed independently. ** Drone-BD ** On top of that, to the best of our knowledge, there are almost none drone-view images for building detection in any public datasets. In order to obtain building proposals with better quality, we also annotate 2,334 building bounding boxes on our 18 drone-view images and 185 images from Dronestagram website for training RPN [5]. This drone-view building detection dataset is called Drone-BD. ** IG-City8 ** Due to the lack of drone-view data, we collect building images based on check-ins (locations) from Instagram and use the hashtag of #building and #buildings in 8 cities including New York, London, Paris, Hong Kong, Tokyo, Sydney, Berlin and San Francisco, to get building images. They are mostly ground-view or street-view images taken by users. We make buildings with the same location be a positive pair and with another random location be a negative pair to form a triplet. We manually remove some noisy images owing to not facing the correct building; eventually, there are 4,409 images, 848 buildings and 104,906 triplets in our Instagram building dataset called IG-City8.

Dataset Attributes

Label SVG
TasksDetection
Label SVG
CategoriesDrone, Building, Spatial Estimation