## Optimal contour closure by superpixel grouping (2010)

Venue: | In ECCV |

Citations: | 11 - 6 self |

### BibTeX

@INPROCEEDINGS{Levinshtein10optimalcontour,

author = {Alex Levinshtein and Cristian Sminchisescu and Sven Dickinson},

title = {Optimal contour closure by superpixel grouping},

booktitle = {In ECCV},

year = {2010}

}

### OpenURL

### Abstract

Abstract. Detecting contour closure, i.e., finding a cycle of disconnected contour fragments that separates an object from its background, is an important problem in perceptual grouping. Searching the entire space of possible groupings is intractable, and previous approaches have adopted powerful perceptual grouping heuristics, such as proximity and co-curvilinearity, to manage the search. We introduce a new formulation of the problem, by transforming the problem of finding cycles of contour fragments to finding subsets of superpixels whose collective boundary has strong edge support in the image. Our cost function, a ratio of a novel learned boundary gap measure to area, promotes spatially coherent sets of superpixels. Moreover, its properties support a global optimization procedure using parametric maxflow. We evaluate our framework by comparing it to two leading contour closure approaches, and find that it yields improved performance. 1

### Citations

449 |
Perceptual organization and visual recognition
- Lowe
- 1999
(Show Context)
Citation Context ...es exploited by such approaches can clearly play a powerful role in detecting closure. Many researchers have exploited the classical Gestalt cues of parallelism and symmetry to group contours. Lowe’s =-=[10]-=- early work on perceptual grouping was one of the first to develop a computational model for parallelism, collinearity, and proximity. Many computational models exist for symmetry-based grouping, incl... |

407 | Learning to detect natural image boundaries using local brightness, color, and texture cues
- Martin, Fowlkes, et al.
- 1994
(Show Context)
Citation Context ...boundaries. Given a contour image I(x, y) 1 , we first segment it into N superpixels using a modified version of the superpixel segmentation method of Mori et al. [25] ([25] uses the Pb edge detector =-=[26]-=-, while we use globalPb [24]). If we let Xi be a binary indicator variable for the i-th superpixel, the vector X yields a full labeling of the superpixels of I as figure (1) or ground (0). Recall that... |

176 | Stochastic completion fields: A neural model of illusory contour shape and salience,” ICCV
- Williams, Jacobs
- 1995
(Show Context)
Citation Context ...-, mid-, and high-level shape priors, as exemplified by Ren et al. [7]. We will also stop short of reviewing methods focused solely on contour completion, e.g., Ren et al. [8] and Williams and Jacobs =-=[9]-=-, although the regularities exploited by such approaches can clearly play a powerful role in detecting closure. Many researchers have exploited the classical Gestalt cues of parallelism and symmetry t... |

172 | Recovering human body configurations: combining segmentation and recognition
- Mori, Ren, et al.
- 2004
(Show Context)
Citation Context ...ting closure to lie along superpixel boundaries. Given a contour image I(x, y) 1 , we first segment it into N superpixels using a modified version of the superpixel segmentation method of Mori et al. =-=[25]-=- ([25] uses the Pb edge detector [26], while we use globalPb [24]). If we let Xi be a binary indicator variable for the i-th superpixel, the vector X yields a full labeling of the superpixels of I as ... |

166 | Class-speci c, top-down segmentation
- Borenstein, Ullman
- 2002
(Show Context)
Citation Context ...rs (RRC) from Stahl and Wang [4]. We provide a qualitative evaluation on various images (see Fig. 6), as well as a quantitative evaluation on two datasets, including the Weizmann Horse Database (WHD) =-=[27]-=- and the Weizmann Segmentation Database (WSD) [28]. Learning the gap measure (Sec. 4) is accomplished on the first 30 images from WHD. For testing, we use 170 additional images from WHD and all 100 im... |

111 |
Smoothed Local Symmetries and Their Implementation
- Brady, Asada
- 1984
(Show Context)
Citation Context ...ual grouping was one of the first to develop a computational model for parallelism, collinearity, and proximity. Many computational models exist for symmetry-based grouping, including Brady and Asada =-=[11]-=-, Cham and Cipolla [12], Saint-Marc et al. [13], Ylä-Jääski and Ade [14], and more recently, Stahl and Wang [15]. One significant challenge faced by these systems is the complexity of pairwise contour... |

99 |
Laws of organization in perceptual forms
- Wertheimer
- 1938
(Show Context)
Citation Context ... background. What makes the problem particularly hard is the intractable number of cycles that may exist in the contours extracted from an image of a real scene. Early perceptual grouping researchers =-=[1]-=- identified a set of nonaccidental contour relations, such as symmetry, parallelism, collinearity, cocurvilinearity, etc., that can be used to link together causally related contours. Such nonaccident... |

85 | Computing contour closure
- Elder, Zucker
- 1996
(Show Context)
Citation Context ...priors, such as continuity and proximity. The most basic closurebased cost function uses a notion of boundary gap, which is a measure of missing image edges along the closed contour. Elder and Zucker =-=[18]-=- model the probability of a connection between two adjacent contour fragments, and find contour cycles using a shortest path algorithm. Wang et al. [19] optimize a measure of average gap using the rat... |

84 |
Using contours to detect and localize junctions in natural images,” CVPR
- Maire, Arbeláez, et al.
- 2008
(Show Context)
Citation Context ...image I(x, y) 1 , we first segment it into N superpixels using a modified version of the superpixel segmentation method of Mori et al. [25] ([25] uses the Pb edge detector [26], while we use globalPb =-=[24]-=-). If we let Xi be a binary indicator variable for the i-th superpixel, the vector X yields a full labeling of the superpixels of I as figure (1) or ground (0). Recall that our goal will be to select ... |

77 |
Robust and efficient detection of salient convex groups
- Jacobs
- 1996
(Show Context)
Citation Context ...uperpixels, but will focus on the more generic cue of closure. Further down the spectrum of prior knowledge are methods based on weaker shape priors than parallelism and symmetry. For example, Jacobs =-=[17]-=- uses convexity as well as gap to extract closed contours by grouping straight line segments. A less restrictive measure is that of compactness, which can be attained by normalizing the gap by area (E... |

72 | Globally optimal regions and boundaries as minimum ratio weight cycles
- Jermyn, Ishikawa
(Show Context)
Citation Context ...rs correspond to circles in that space; however, their goal is to better detect object contours, and they stop short of grouping the contours into closed boundaries. The method of Jermyn and Ishikawa =-=[23]-=- is perhaps the closest to our work. Similar to [19,4], they minimize closure costs using ratio cuts, but unlike [19,4] who operate on contour fragments, [23] works directly with pixels in a 4-connect... |

60 | Constrained Parametric Min-Cuts for Automatic Object Segmentation - Carreira, Sminchisescu - 2010 |

49 |
G.: B-spline contour representation and symmetry detection
- Saint-Marc, Rom, et al.
- 1993
(Show Context)
Citation Context ...computational model for parallelism, collinearity, and proximity. Many computational models exist for symmetry-based grouping, including Brady and Asada [11], Cham and Cipolla [12], Saint-Marc et al. =-=[13]-=-, Ylä-Jääski and Ade [14], and more recently, Stahl and Wang [15]. One significant challenge faced by these systems is the complexity of pairwise contour grouping to detect symmetry-related contour pa... |

40 | Cue integration in figure/ground labeling
- Ren, Fowlkes, et al.
- 2007
(Show Context)
Citation Context ...ill see, many make assumptions about the nature of parts that make up the objects in the scene. In fact, some methods incorporate low-, mid-, and high-level shape priors, as exemplified by Ren et al. =-=[7]-=-. We will also stop short of reviewing methods focused solely on contour completion, e.g., Ren et al. [8] and Williams and Jacobs [9], although the regularities exploited by such approaches can clearl... |

40 | Untangling cycles for contour grouping
- Zhu, Song, et al.
(Show Context)
Citation Context ...exity, they either resort to heuristics to prune the search (e.g., [17]) or constrain the search space by other means (e.g., restricting the closure to alternating gap/non-gap cycles [4]). Zhu et al. =-=[22]-=- propose to solve this hard grouping problem by embedding the edge fragments into polar coordinates such that closed contours correspond to circles in that space; however, their goal is to better dete... |

39 | Applications of parametric maxflow in computer vision
- Kolmogorov, Boykov, et al.
- 2007
(Show Context)
Citation Context ...wo components of our cost function, i.e., area and gap, are combined in a simple ratio that can be efficiently optimized482 A. Levinshtein, C. Sminchisescu, and S. Dickinson using parametric maxflow =-=[5]-=- to yield the global optimum. The optimal solution yields the largest set of superpixels bounded by contours that have the least gaps (Fig. 1(d)). Moreover, parametric maxflow will yield the top k sol... |

38 | Salient closed boundary extraction with ratio contour,” PAMI
- Wang, Kubota, et al.
- 2005
(Show Context)
Citation Context ...dges along the closed contour. Elder and Zucker [18] model the probability of a connection between two adjacent contour fragments, and find contour cycles using a shortest path algorithm. Wang et al. =-=[19]-=- optimize a measure of average gap using the ratio cut approach. However, a measure based purely on the total boundary gap is insufficient for perceptual closure, and Elder and Zucker [20] argue that ... |

32 | Scale-invariant contour completion using conditional random fields,” ICCV
- Ren, Fowlkes, et al.
- 2005
(Show Context)
Citation Context ...some methods incorporate low-, mid-, and high-level shape priors, as exemplified by Ren et al. [7]. We will also stop short of reviewing methods focused solely on contour completion, e.g., Ren et al. =-=[8]-=- and Williams and Jacobs [9], although the regularities exploited by such approaches can clearly play a powerful role in detecting closure. Many researchers have exploited the classical Gestalt cues o... |

31 | A.R.: Perceptual completion of occluded surfaces
- Williams, Hanson
- 1996
(Show Context)
Citation Context ...ed purely on the total boundary gap is insufficient for perceptual closure, and Elder and Zucker [20] argue that the distribution of gaps along the contour is also very important. Williams and Hanson =-=[21]-=- addressed the problem of perceptual completion of occluded surfaces, formulated as the problem of computing a labeled knotdiagram representing a set of occluded surfaces from observed image contours.... |

27 | Image segmentation by probabilistic bottom-up aggregation and cue integration,” CVPR
- Alpert, Galun, et al.
- 2007
(Show Context)
Citation Context ...litative evaluation on various images (see Fig. 6), as well as a quantitative evaluation on two datasets, including the Weizmann Horse Database (WHD) [27] and the Weizmann Segmentation Database (WSD) =-=[28]-=-. Learning the gap measure (Sec. 4) is accomplished on the first 30 images from WHD. For testing, we use 170 additional images from WHD and all 100 images from WSD. 6.1 Quantitative Evaluation For a q... |

21 | R.: Geometric saliency of curve correspondances and grouping of symmetric comntours
- Cham, Cipolla
- 1996
(Show Context)
Citation Context ... the first to develop a computational model for parallelism, collinearity, and proximity. Many computational models exist for symmetry-based grouping, including Brady and Asada [11], Cham and Cipolla =-=[12]-=-, Saint-Marc et al. [13], Ylä-Jääski and Ade [14], and more recently, Stahl and Wang [15]. One significant challenge faced by these systems is the complexity of pairwise contour grouping to detect sym... |

20 | A measure of closure
- ELDER, ZUCKER
- 1994
(Show Context)
Citation Context ...Wang et al. [19] optimize a measure of average gap using the ratio cut approach. However, a measure based purely on the total boundary gap is insufficient for perceptual closure, and Elder and Zucker =-=[20]-=- argue that the distribution of gaps along the contour is also very important. Williams and Hanson [21] addressed the problem of perceptual completion of occluded surfaces, formulated as the problem o... |

16 | F.: Grouping symmetrical structures for object segmentation and description
- Ylä-Jääski, Ade
- 1996
(Show Context)
Citation Context ...arallelism, collinearity, and proximity. Many computational models exist for symmetry-based grouping, including Brady and Asada [11], Cham and Cipolla [12], Saint-Marc et al. [13], Ylä-Jääski and Ade =-=[14]-=-, and more recently, Stahl and Wang [15]. One significant challenge faced by these systems is the complexity of pairwise contour grouping to detect symmetry-related contour pairs. Levinshtein et al. [... |

14 | A.D.: Perceptual grouping for contour extraction
- Estrada, Jepson
(Show Context)
Citation Context ...elated contours. Such nonaccidental grouping rules can serve as powerful heuristics to help manage the complexity of greedily searching for a contour closure that is unlikely to have arisen by chance =-=[2,3]-=-. However, the space of possible closures is still overwhelming, particularly when one allows larger and larger boundary gaps in a closure. Finding an optimal solution is intractable without somehow r... |

12 | Multiscale symmetric part detection and grouping
- Levinshtein, Sminchisescu, et al.
- 2009
(Show Context)
Citation Context ...], and more recently, Stahl and Wang [15]. One significant challenge faced by these systems is the complexity of pairwise contour grouping to detect symmetry-related contour pairs. Levinshtein et al. =-=[16]-=- attempt to overcome this computational complexity limitation by constraining theOptimal Contour Closure by Superpixel Grouping 483 symmetric parts to be collections of superpixels. We will draw on t... |

7 | S.: Edge grouping combining boundary and region information - Stahl, Wang |

7 | Globally optimal grouping for symmetric closed boundaries by combining boundary and region information
- Stahl, Wang
(Show Context)
Citation Context ... Many computational models exist for symmetry-based grouping, including Brady and Asada [11], Cham and Cipolla [12], Saint-Marc et al. [13], Ylä-Jääski and Ade [14], and more recently, Stahl and Wang =-=[15]-=-. One significant challenge faced by these systems is the complexity of pairwise contour grouping to detect symmetry-related contour pairs. Levinshtein et al. [16] attempt to overcome this computation... |

5 | Robust boundary detection with adaptive grouping
- Estrada, Jepson
(Show Context)
Citation Context ...elated contours. Such nonaccidental grouping rules can serve as powerful heuristics to help manage the complexity of greedily searching for a contour closure that is unlikely to have arisen by chance =-=[2,3]-=-. However, the space of possible closures is still overwhelming, particularly when one allows larger and larger boundary gaps in a closure. Finding an optimal solution is intractable without somehow r... |

1 |
S.: A measure of closure. Vision Research 34
- Elder, Zucker
- 1994
(Show Context)
Citation Context ...Wang et al. [19] optimize a measure of average gap using the ratio cut approach. However, a measure based purely on the total boundary gap is insufficient for perceptual closure, and Elder and Zucker =-=[20]-=- argue that the distribution of gaps along the contour is also very important. Williams and Hanson [21] addressed the problem of perceptual completion of occluded surfaces, formulated as the problem o... |