REU Presentation Week 3 Nicholas Baker
What features “pop out” in a scene? No prior information/goal Identify areas of large feature contrasts in center-surround condition Luminance, color, orientation, motion Bottom Up Visual Salience
Identify areas of high intrinsic dimensionality by analyzing the signal as Shannon information (Vig 2012) Identify areas of low level surprisal in a scene (Itti 2005) Weight continuity and visual clutter as well as local feature contrasts (He 2011) Separate feature matrix into low rank non-salient matrix and sparse salient matrix (Souly) Bottom up Visual Salience in Computer Vision
Goal driven analysis of scene Direct visual attention to area/features of probable importance Locate objects/actions/features of exogenous significance Top Down Visual Salience
Use CRF modulated dictionary learning to construct top down saliency map (Yang 2012) Use online Reinforced Learning to interactively teach machine how to correctly allocate attention using U-Tree algorithm (Borji 2009) Top Down Visual Salience in Computer Vision
Most current top-down visual saliency work is on static images Choose one promising top-down method for static images Implement the algorithm if code is not available Extend it to perform on videos instead of static images My Work