Automatic Retrieval of Near Photo-Realistic Textures from Single Ground-Level Building Images

Abstract

An approach is presented for the automatic retrieval of near photo-realistic textures from single ground-level building images. First, the building facade texture is extracted using the Watershed segmentation technique, which is carried out repetitively until the most successful segment is extracted. Next, the retrieved building image is geometrically rectified in an automated way. After that the occlusions are removed using an image-based approach, which includes marking the occluded region, searching the marked region in the remaining parts of the facade image, selecting a candidate patch with the highest correlation, and copying it into the marked region. The developed concept was tested using two distinct data sets. The first data set contains fifteen rectilinear buildings selected from the residential area of the Batikent district of Ankara, Turkey. The second dataset includes five images selected from eTRIMS database, which contains over a hundred buildings captured in major European cities. The qualitative results obtained are quite promising. For the assessment of facade image extraction, both datasets provided a quantitative accuracy of above 80%. In the rectification results, 15 out of 20 buildings produced the positional mean errors of below 9 pixels. The subjective assessment of the occlusion removal yielded the mean rating scores of 2.58 and 2.28 for the Batikent and eTrims datasets, respectively. While the rating score of 2.58 can be categorized in the middle of the criterions "Fine" and "Passable", the score 2.28 would be "Fine".