Arithmetically intensive operations, replicated over huge data sets (usually image pixels or scanned data), are an important part of many Computer Vision (CV) tasks, making them good candidates for taking advantage of the processing power of contemporary graphics processing units (GPUs). This thesis formulates a set of CV algorithms that use low level representations of visual content and are tailored for running on GPUs. A general view of GPUs and parallel programming patterns that offers interesting building blocks for CV tasks provides the necessary background for the algorithms. We also propose a formal definition for the Multiple Reduction pattern and evaluate its efficiency according to different reduction factors and layouts. We present two techniques for extracting data from the image space using the GPU: MOCT, a technique for tracking a set of objects identified by their colors from natural videos, and MRR, a technique for distributing the evaluation of a set of operators defined over different regions of interest within an image. As a MRR application we describe a Centroidal Voronoi Diagram construction based on Lloyds algorithm but entirely computed using GPU resources. We also deal with visual content representations as pixel agglomerations, more specifically, as Regional Quadtrees. We introduce the QuadN4tree: a new model for representing quadtree leaves that allows navigation through their neighborhood systems and achieves an optimal cost for the retrieval of neighbor sets. We also propose modifying the setup for CV applications based on energy minimization via graph cuts, introducing a preprocessing step that groups similar pixels into regional quadtree leaves. This modification aims to reduce the size of the graph for which a minimum cut is to be found. We apply our proposed method to the problem of natural image segmentation by active illumination. Published papers detailing some contributions of this dissertation are included as appendixes. They present data-parallel formulations for the CV tasks we describe.