A Novel Dataset of Structured Probability Distributions in Natural Scenes
Abstract
The study of natural scene statistics has served as a powerful framework for understanding vision and neural coding in the last several decades. Critical to this framework are datasets of natural scenes that have aligned multi-modal visual information, including luminance, color, stereoscopic disparity, movement, and three-dimensional (3D) information, which we are acquiring with support of DURIP grants from ARL/ARO and AFOSR. With support of this STIR grant, we performed statistical analyses on these datasets and developed a set of probabilistic models, referred to as probabilistic visual codes (PVCs). The PVCs are probabilistic models of static and dynamic, 2D and 3D natural scene patches in center-surround configurations. We found that these PVCs have a universal geometry: each PVC is a function of the total distance to hyperplanes in the spaces of 2D and/or 3D visual features in space and/or time domains and a large set of hyperplanes partition the feature spaces so that any natural scene patch is a combination of samples of PVCs. We are now exploring ways to relate PVCs to neural encoding and visual learning and applications of PVCs to visual saliency, natural 3D vision, scene vision, visual memory, object perception, and dynamic scene understanding.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jul 30, 2015
- Accession Number
- AD1001049
Entities
People
- Zhiyong Yang
Organizations
- Medical College of Georgia