Voxel-based 3D Model Reconstruction from Images and Silhouettes

Link to full Thesis

In this thesis, we propose a 3d model reconstruction system from images and silhouettes, and its output models are printable. We also studied characteristics and problems of images shot with mobile phones, and altered an existing method in order to reconstruct 3d model robustly.

We built a multi-view stereo system based on Volumetric Graph-cuts, which generates models composed of voxels. The structure of the system is described by an energy function, and a Markov Random Field is built to optimize. Finally, all the voxels in the space are divided into model and background.

Because of images captured with model phones, noise and artifacts are hardly avoided. However, existing multi-view stereo methods are not noise-tolerant and thus produce bad models impacted by noise. Hence, this thesis digged into the interference of noise, and came up with two solutions: (1) Shrinking images before computation, and (2) Add threshold filter to the original algorithm.

In conclusion, in the thesis we combined the photo clue from color images and shape constraint from silhouettes into a system. The system not only accepts noisy images but also produces printable models.