CV-NICS: a lightweight solution to the correspondence problem

In this dissertation, I present a novel approach for solving the correspondence problem using basic statistical classification techniques. While metrics such as Pearson's rho or cosine similarity would not be powerful enough to solve the correspondence problem directly, their performance can be enhanced by augmenting the scene with random color static via a projector. Over time, this noise increases the statistical independence of imaged points not in correspondence. This allows the reduction of the correspondence problem to a simple similarity search of temporal features. Extensive experiments have shown the approach to be as effective as more complex structured light techniques at producing very dense correspondence data for a variety of scenes. The approach differentiates itself from traditional structured lighting by not relying on known camera or projector geometries, and by allowing relatively lax capturing conditions. Due to the statistically oriented nature of the approach and unlike more recognition focused techniques, the approach is naturally amenable to quality assessment and analysis. This dissertation provides a background on the correspondence problem, presents empirical and analytical results regarding the new technique, and reviews the related work in the literature.