Browsing by Author "Väänänen, Pekka"

Real-world locations can be reconstructed as digital 3D models using 3D scanning. The scans often suffer from missing surface regions caused by occlusions and poor scanning geometry, limiting their usefulness for many tasks. We present an automated system that repairs small missing regions using generative adversarial networks (GANs). The system operates on heightmaps of small round surface patches distributed around the missing region. A neural network model predicts a complete plausible surface for each corrupted patch, which is then integrated to the scan. In addition to geometry, surface colors are also generated. Encouraging results are found in the color reconstruction task, but the output geometry is not clearly superior to the results of a simpler baseline spline model.