University of Twente Student Theses


Semantic building facade segmentation from airborne oblique images

Lin, Yaping (2018) Semantic building facade segmentation from airborne oblique images.

[img] PDF
Abstract:With the introduction of airborne oblique camera systems and the improvement of photogrammetric techniques, high-resolution2D and 3D data can be acquired in urbanareas. This high-resolution data allows us toperform detailed investigations about building roofs and façades which can contribute to LoD3 city modelling. Normally,façade segmentation is preformedfrom terrestrial views. However, acquiring terrestrial images is time-consuming when it comes to largeurban areas. In my study, high resolution aerial oblique images are used as data source for façade segmentation in urbanareas.In addition to traditional image features, like RGB and SIFT, normal vector and planarity are also extractedat different scalesfrom densematching point cloud. Then, these 3D geometrical features are projected back to 2D space to assist façade interpretation. Random forest is trained and applied to label façade pixels. Outputs of random forest are always noisy because no contextual information is taken into consideration. As a result, conditional random field (CRF) is applied to refine classification results by involving neighboring information.The experiment is conducted in three different scenarios where different training strategies areused, 3-class classification, 5-class classification and 5-class-equal classification.In all scenarios, three CRF models are implemented, namely 8 connected CRF, higher order CRF and fully connected CRF. In 8 connected CRF, for each pixel, it only connects to its 8 nearest pixels. This takes very limited contextual information. Therefore, another potential term, computed based on superpixels got from unsupervised segmentationbased on surface growing algorithm in 3D space, is added to 8 connected CRF to enforce label consistency. This is called as higher order CRF.Fully connected CRF, connecting all pixels in pairs, allows to captureglobal interactions over an image. 8 connected and higher order CRFs are solved by graph cut based algorithms while fully connected CRF is solved by mean field approximation.Experiments show that adding 3D features can significantly improve classification IoUby 26.36%, 15.57% and 11.92% respectively in 3-class, 5-class and 5-class-equal scenarios. In terms of CRF models, fully connected potentialsperform the best in refining resultsin 3-class and 5-class-equal scenarios, improving IoUby 4.67% and 6.75%
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page