GEO-TAGGED IMAGE RETRIEVAL FROM MAPILLARY STREET IMAGES FOR A TARGET BUILDING
This study aims to investigate the possibility to automate the image selection process for the target building from Mapillary images through a web application where the user only initiates one image of the target building as a query. Using the data provided with Mapillary API and Overpass API, all images having full or partial coverage of the target building were selected. Then the images were segmented by using a pre-trained U-Net model to discard any images having less than 20% building coverage. The experiments showed promising results yielding 0.971 and 0.887 of overall accuracy after segmentation steps for two different target buildings.