Reflectance describes how a material reflects light. Obtaining the reflectance of real materials has many applications, from virtual asset creation to coherent illumination in mixed reality. Due to lower cost and less required expertise it would be useful to obtain reflectance from images instead of specialized measuring equipment. This is an inherently ill posed problem because pixel intensity is dependant on many ambiguities, such as the strength of incident light and reflectivity of surfaces. A large amount existing work has attempted to overcome these ambiguities, initially with statistical methods and more recently with deep learning. However, the task remains difficult for unconstrained images. The case of casually captured images of complex real world scenes has been insufficiently addressed. The aim of this thesis is to develop a method for accurately estimating a parametric description of reflectance for every pixel in such images. To achieve this, deep learning methods will be combined with sensors and estimators from the related fields of geometry and light capture. In this proposal, research objectives are established and justified, preliminary work and results indicating that an effective methodology has been chosen is presented, and a concrete plan for continued research is made.