3D Reconstruction using Prior Knowledge

Zusammenfassung

This dissertation investigates how prior knowledge can improve optical reconstruction systems that enable the localizationand measurement of objects in three-dimensional space. Reconstruction algorithms are usually subject to information lossduring the initial image acquisition, affecting the results’ quality. This fundamental problem can be mitigated by exploiting prior knowledge about the scene. Four reconstructionsystems are presented in this dissertation, demonstrating theeffective utilization of prior knowledge. First, a deep convolutional neural network is used for matting dynamic scenes, exploiting synchronizedbackground color changes to reconstruct transparent foregrounds, even with imprecise backgroundcolors. Second, the exact positions of cochlear implant electrodes are localized using Markov random fields, utilizing priorknowledge of electrode distances and minimal bending radii, significantly improving positioning accuracy. Third, electroluminescent wires woven into hair help reconstruct braided hairstyles for special effects by using active curves to track guidehairs and create realistic 3D braid models. Finally, 3D reconstruction of the human spine during movement is ...