GrabCut is an iterative process that in cutting out foreground object after user specify the approximate region(rectangle) surrounding it. It uses K components of Gaussian Mixture Models to model the color of foreground and background pixels separately. The Graph-Cut technique is used to classify pixels as background / foreground. The factors affecting the classification is the neighborhood color gradients and fitness into existing GMM foreground/background models. The iteration begins with updating the 2K GMMs from the newly classified pixels (Learning), followed by detecting the foreground pixels with Graph Cut.
Sample
- OpenCV supports 4 types of classifications: BG, FG, Likely-BG and Likely-FG in user-edit mode.
- FG pixels are classified by GrabCut as Likely-FG, not FG (by experiment).
- Require manual iteration to observe convergence.
- It looks like the iterations are trying to make better GMMs (scissors.jpg).
- person1.jpg takes 5 seconds on each iteration; after 6 iterations there is no much improvements.
- Hard to get the narrow 'stem' classified as FG after putting a rectangle around the plant (besides the pot) from bush.jpg. Is it because the relatively small area to accummulate a significant weight in GMMs?
Readings
No comments:
Post a Comment