Scale-Invariant Object Contour Points (SIOCP): Integrating Instance Segmentation, Computer Vision and Geometric Refinement for Exact Contour and Keypoint Extraction of Façade Elements from Urban Images

Abstract. Exact 6DoF pose estimation is key to achieving high-quality CityGML LoD3 and BIM façade element reconstruction from urban monocular 2D RGB images. We present an implementation and experimental evaluation of our extraction pipeline for Scale- Invariant Object Contour Points (SIOCP). Our pipeline fuses processed information from 2D RGB images, IMU, GNSS+RTK, and LoD2 data to derive precise contours of regularly shaped objects and stable keypoints for 6DoF pose estimation. The image-processing pipeline is implemented using YOLOv8, SAM, and enhanced classical algorithms. Furthermore, the keypoint descriptor is based on a hierarchical object catalog that fully describes façades and the interdependencies of their elements in spatial and temporal contexts. This descriptor methodology enables reliable keypoint matching across texture-rich and detailed façade images from different perspectives, where classical methods partially fail in the benchmark. The implementation is still being refined. In summary, we present SIOCP as a key component for accurate 6DoF pose reconstruction, with future integration planned for LoD3 and beyond in building reconstruction.