Chapter 6 – Anchored Regression

This chapter first reviews the general framework for anchored regression, based on a relaxation of the sparsity constraint in sparse coding that enables training linear regressors for each atom in the optimized dictionary. Then the chapter proceeds with exploring two extensions toward better quality, based on training with all example patches, rather than using just the optimized dictionary atoms, and improved runtime, for example, through state-of-the-art hashing techniques. The incorporation of the latter further pushes the usability of anchored regression to a broader set of applications with tighter temporal constraints.