## Optimized Color Sampling for Robust Matting

### Cached

### Download Links

- [juew.org]
- [research.microsoft.com]
- [www.research.microsoft.com]
- [grail.cs.washington.edu]
- [research.microsoft.com]
- [research.microsoft.com]
- DBLP

### Other Repositories/Bibliography

Citations: | 48 - 3 self |

### BibTeX

@MISC{Wang_optimizedcolor,

author = {Jue Wang},

title = {Optimized Color Sampling for Robust Matting},

year = {}

}

### OpenURL

### Abstract

Image matting is the problem of determining for each pixel in an image whether it is foreground, background, or the mixing parameter, ”alpha”, for those pixels that are a mixture of foreground and background. Matting is inherently an ill-posed problem. Previous matting approaches either use naive color sampling methods to estimate foreground and background colors for unknown pixels, or use propagation-based methods to avoid color sampling under weak assumptions about image statistics. We argue that neither method itself is enough to generate good results for complex natural images. We analyze the weaknesses of previous matting approaches, and propose a new robust matting algorithm. In our approach we also sample foreground and background colors for unknown pixels, but more importantly, analyze the confidence of these samples. Only high confidence samples are chosen to contribute to the matting energy function which is minimized by a Random Walk. The energy function we define also contains a neighborhood term to enforce the smoothness of the matte. To validate the approach, we present an extensive and quantitative comparison between our algorithm and a number of previous approaches in hopes of providing a benchmark for future matting research. 1.

### Citations

703 | Grabcut-interactive foreground extraction using iterated graph cuts, in
- Rother, Kolmogorov, et al.
(Show Context)
Citation Context ...cially when the confidence value is high; secondly, the matte should be locally smooth and robust to image noise (neighborhood constraint). As shown in previous graph-based image labelling approaches =-=[14, 9]-=-, this expectation can be satisfied by solving a graph labelling problem shown in Figure 6, where ΩF and ΩB are virtual nodes representing pure foreground and pure background, white nodes represent un... |

237 | Random walks for image segmentation
- Grady
- 2006
(Show Context)
Citation Context ...3.1 Solving for Optimal αs Given the fact that alpha values are continuous, we avoid discrete labelling optimizations such as graph cut. Instead, we solve the graph labelling problem as a Random Walk =-=[3, 4]-=-, which has been shown to minimize the total graph energy over real values. Although a detailed description of Random Walk theory is beyond the scope of this paper, it essentially works as follows. Fi... |

195 | A Bayesian approach to digital matting
- Chuang, Curless, et al.
- 2001
(Show Context)
Citation Context ...on to estimate alphas. Ruzon and Tomasi [10] analyze the statistical distributions of foreground and background samples for alpha estimation. Their approach is improved by the Bayesian matting system =-=[1]-=-, which formulates the problem in a well-defined Bayesian framework and solves it using the MAP technique. The recently proposed Belief Propagation matting system [14] and easy matting system [6] solv... |

131 |
Blue screen matting
- Smith, Blinn
- 1996
(Show Context)
Citation Context ...n be minimized globally. 2.3. Extensions Instead of a single input image, additional information can be used for reducing unknowns in matte estimation if available, such as using multiple backgrounds =-=[11]-=-, or flash and non-flash image pairs [13]. In this paper we only consider the general problem of matte estimation from a single image with limited user input. 3. Failure Modes for Previous Approaches ... |

103 | Alpha estimation in natural images
- Ruzon, Tomasi
- 2000
(Show Context)
Citation Context ...image matting was first proposed in [8]. Following this idea, the KnockOut2 system [2] extrapolates known foreground and background colors into the unknown region to estimate alphas. Ruzon and Tomasi =-=[10]-=- analyze the statistical distributions of foreground and background samples for alpha estimation. Their approach is improved by the Bayesian matting system [1], which formulates the problem in a well-... |

95 | Poisson matting
- Sun, Jia, et al.
- 2004
(Show Context)
Citation Context ...nd and background estimation. 2.2. Propagation-based Approaches Propagation-based approaches solve the matte without explicit foreground and background color estimation. The Poisson matting algorithm =-=[12]-=- assumes the foreground and background colors are smooth in a narrow band of unknown pixels. Thus, the gradient of the matte matches with the gradient of the image, which can be calculated by solving ... |

91 | An iterative optimization approach for unified image segmentation and matting
- Wang, Cohen
- 2005
(Show Context)
Citation Context ...ed by the Bayesian matting system [1], which formulates the problem in a well-defined Bayesian framework and solves it using the MAP technique. The recently proposed Belief Propagation matting system =-=[14]-=- and easy matting system [6] solve a matte directly from a few user specified scribbles instead of a carefully specified trimap. Since the user input is very sparse in this case, global sampling metho... |

90 | A closed form solution to natural image matting
- Levin, Lischinski, et al.
- 2006
(Show Context)
Citation Context ... gradient of the matte matches with the gradient of the image, which can be calculated by solving Poisson equations. A similar method based on Random Walks is proposed in [5]. The Closed-form matting =-=[7]-=- approach assumes foreground and background colors can be fit with linear models in local windows, which leads to a quadratic cost function in alpha that can be minimized globally. 2.3. Extensions Ins... |

55 | Knockout user guide
- CORPORATION
- 2002
(Show Context)
Citation Context ...ampling-based Approaches Sampling representative foreground and background colors and analyzing their statistics for image matting was first proposed in [8]. Following this idea, the KnockOut2 system =-=[2]-=- extrapolates known foreground and background colors into the unknown region to estimate alphas. Ruzon and Tomasi [10] analyze the statistical distributions of foreground and background samples for al... |

37 | Random walks for interactive alpha-matting
- Grady, Schiwietz, et al.
- 2005
(Show Context)
Citation Context ... of unknown pixels. Thus, the gradient of the matte matches with the gradient of the image, which can be calculated by solving Poisson equations. A similar method based on Random Walks is proposed in =-=[5]-=-. The Closed-form matting [7] approach assumes foreground and background colors can be fit with linear models in local windows, which leads to a quadratic cost function in alpha that can be minimized ... |

15 | Flash matting
- Sun, Li, et al.
- 2006
(Show Context)
Citation Context ...Instead of a single input image, additional information can be used for reducing unknowns in matte estimation if available, such as using multiple backgrounds [11], or flash and non-flash image pairs =-=[13]-=-. In this paper we only consider the general problem of matte estimation from a single image with limited user input. 3. Failure Modes for Previous Approaches Before introducing the proposed robust ma... |

12 | Soft edge chroma-key generation based upon hexoctahedral color space - Mishima - 1993 |

5 |
Easy matting
- Guan, Chen, et al.
- 2006
(Show Context)
Citation Context ...stem [1], which formulates the problem in a well-defined Bayesian framework and solves it using the MAP technique. The recently proposed Belief Propagation matting system [14] and easy matting system =-=[6]-=- solve a matte directly from a few user specified scribbles instead of a carefully specified trimap. Since the user input is very sparse in this case, global sampling methods are proposed to assist th... |

2 |
Multilabel randomwalker image segmentation using prior models
- Grady
- 2005
(Show Context)
Citation Context ...3.1 Solving for Optimal αs Given the fact that alpha values are continuous, we avoid discrete labelling optimizations such as graph cut. Instead, we solve the graph labelling problem as a Random Walk =-=[3, 4]-=-, which has been shown to minimize the total graph energy over real values. Although a detailed description of Random Walk theory is beyond the scope of this paper, it essentially works as follows. Fi... |