AUTOMATING GRAB-CUT FOR SINGLE- OBJECT FOREGROUND IMAGES Eugene Weiss Computer Vision Stanford University December 14, 2011 Eugene Weiss 1
Automating Grab-Cut for Single-Object Foreground Images Grab-Cut: Interactive Foreground Extraction Using Iterated Graph Cuts. Invented by Rother, Kolmogorov and Blake at Microsoft Research, Cambridge in User defines a rectangular bounding box to separate rough foreground from background. Eugene Weiss
Automating Grab-Cut for Single-Object Foreground Images Automating Grab Cut A Good bounding box doesn’t need to frame the image perfectly. Better thought of as background subtraction than foreground identification. So it’s better to think “outside the box”. Eugene Weiss
Automating Grab-Cut for Single-Object Foreground Images Eugene Weiss
Automating Grab-Cut for Single-Object Foreground Images First (Naïve) Pass The area in red represents the tested background slice, the rest is treated as foreground. Eugene Weiss
Automating Grab-Cut for Single-Object Foreground Images Second Pass 3 sides of the naïve box are used for foreground, the slice is constrained by the perpendicular edges of the box. Eugene Weiss
Automating Grab-Cut for Single-Object Foreground Images An Easy Case A clear color differentiation of the foreground from the background makes for a good result. Eugene Weiss
Automating Grab-Cut for Single-Object Foreground Images A good result with a challenging image A good outcome despite a similar background color. Eugene Weiss
Automating Grab-Cut for Single-Object Foreground Images Problems Images intersecting the image border present a challenge. Eugene Weiss
Automating Grab-Cut for Single-Object Foreground Images Problems A portrait may want a face, or a full torso. Presently, the face is selected. Eugene Weiss
Automating Grab-Cut for Single-Object Foreground Images The Hardest Image. The hardest case is where the foreground object is both complex and intersects the image edge. Eugene Weiss
Automating Grab-Cut for Single-Object Foreground Images Future Directions More features, such as better texture determinants. Look at segmenting the foreground area to avoid cropping. Use threshold to broaden the box, and hopefully solve the edge intersection prob. Leverage Grab-Cut improvements. Eugene Weiss