#### DMCA

## Superpixels and supervoxels in an energy optimization framework (2010)

Venue: | In ECCV |

Citations: | 44 - 2 self |

### Citations

3782 | Normalized cuts and image segmentation - Shi, Malik - 2000 |

3282 | Rapid object detection using a boosted cascade of simple features
- Viola, Jones
- 2001
(Show Context)
Citation Context ...nce the total number of primitives is greatly reduced [9]. Computational efficiency also comes from a reduction in the number of hypothesis. Instead of exhaustive examining of all rectangular patches =-=[10]-=-, an alternative is2 Superpixels and Supervoxels in an Energy Optimization Framework to examine only superpixels [1, 2, 4–7]. In addition to efficiency, superpixels are used for computing features th... |

2394 | Mean-Shift: A robust approach toward feature space analysis
- Comaniciu, Meer
(Show Context)
Citation Context ...1, 2, 4–7]. In addition to efficiency, superpixels are used for computing features that need spatial support [3]. To obtain superpixels, one often uses image segmentation algorithms such as meanshift =-=[11]-=-, graph based [12], normalized cuts [13]. To increase the chance that superpixels do not cross object boundaries, a segmentation algorithm is run in an oversegmentation mode. However, most segmentatio... |

1747 | Additive logistic regression: a statistical view of boosting,” The
- Friedman, Hastie, et al.
- 2000
(Show Context)
Citation Context ...g data. From each box/superpixel, we extract features similar to those used in [3]. We use features based on color, position (relative to the image size) in the image, and texture. We use Gentleboost =-=[29]-=- for training 1 . The testing error is as follows. Our Compact superpixels: 20.5%, FH superpixels [12]: 27.4%, rectangular boxes: 24.0%. Thus performance with our superpixels is significantly better t... |

1312 | An experimental comparison of min-cut/max-flow algorithms for energy minimization in vision,”
- Boykov, Kolmogorov
- 2004
(Show Context)
Citation Context ...o coincide with intensity edges. This energy is NP-hard to optimize. We use the expansion algorithm from [18], which guarantees a factor of 2 approximation. For the max-flow/min-cut algorithm, we use =-=[22]-=-. 2.2 Compact Superpixels First recall the intuitive explanation, Fig. 2. We cover an image with overlapping square patches of fixed size, equal to the maximum allowed superpixel size. WeSuperpixels ... |

953 | A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics
- Martin, Fowlkes, et al.
- 2001
(Show Context)
Citation Context ...n on label l needs to be performed only for pixels in S(l). This is both memory and time efficient. We run the expansion algorithm for two iterations, and it takes about 5 seconds for Berkeley images =-=[15]-=-. Our algorithm would be easy to implement on multiple processors or GPU. To summarize, the properties of compact superpixels are as follows. In the presence of image gradient, superpixel boundaries a... |

939 | Efficient graph-based image segmentation
- Felzenszwalb, Huttenlocher
- 2004
(Show Context)
Citation Context ...ition to efficiency, superpixels are used for computing features that need spatial support [3]. To obtain superpixels, one often uses image segmentation algorithms such as meanshift [11], graph based =-=[12]-=-, normalized cuts [13]. To increase the chance that superpixels do not cross object boundaries, a segmentation algorithm is run in an oversegmentation mode. However, most segmentation algorithms produ... |

489 | BOBICK A.: Graphcut textures: Image and video synthesis using graph cuts.
- KWATRA, Schodl, et al.
- 2003
(Show Context)
Citation Context ...ach to compute superpixels in an energy minimization framework. Our method is simple to understand and implement. The basic algorithm, illustrated in Fig. 2, is similar in spirit to texture synthesis =-=[16]-=-.Superpixels and Supervoxels in an Energy Optimization Framework 3 Fig. 2. Overview of our algorithm. Left: the original image overlayed with square patches. Each patch corresponds to a label. This l... |

415 | C.: A comparative study of energy minimization methods for Markov random fields with smoothness-based priors.
- Szeliski, Zabih, et al.
- 2008
(Show Context)
Citation Context ...other advantage is optimization. Turbopixels are based on level set evolution [17], which is known to have numerical stability issues. We optimize with graph cuts [18], which is known to perform well =-=[19]-=-. Our running time is better. Last, but not least, our approach naturally transfers to 3D for “supervoxel” segmentation of video. An interesting work on superpixels is in [20, 21]. Their goal is somew... |

307 | Graph cuts and efficient nd image segmentation.
- Boykov, Funka-Lea
- 2006
(Show Context)
Citation Context ...these issues after the energy function is completely specified. We now discuss the smoothness term. To better approximate Euclidean metric [23] we use 8-connected N . Vpq is Potts model with wpq from =-=[24]-=-: wpq = exp(− (Ip−Iq)2 dist(p,q)·2σ2 ). Here Ip is the intensity of pixel p, and dist(p, q) is the Euclidean distance between p and q. Observe that with Dp as defined in Eq. (2), the data term in Eq. ... |

281 | Learning a classification model for segmentation
- Ren, Malik
- 2003
(Show Context)
Citation Context ...that are similar in color, texture, etc., and therefore are likely to belong to the same physical world object. The atomic region notion is old, but a popular term superpixel has been coined recently =-=[1]-=-. The assumption that all pixels in a superpixel belong to the same object leads to the advantage of superpixel primitives over pixel primitives. The first advantage is computational efficiency. If on... |

250 | Geometric context from a single image.
- HOIEM, EFROS, et al.
- 2005
(Show Context)
Citation Context ...perpixels and Supervoxels in an Energy Optimization Framework to examine only superpixels [1, 2, 4–7]. In addition to efficiency, superpixels are used for computing features that need spatial support =-=[3]-=-. To obtain superpixels, one often uses image segmentation algorithms such as meanshift [11], graph based [12], normalized cuts [13]. To increase the chance that superpixels do not cross object bounda... |

250 | Computing geodesics and minimal surfaces via graph cuts
- Boykov, Kolmogorov
(Show Context)
Citation Context ...s to use and how to spread them out in the image. We address these issues after the energy function is completely specified. We now discuss the smoothness term. To better approximate Euclidean metric =-=[23]-=- we use 8-connected N . Vpq is Potts model with wpq from [24]: wpq = exp(− (Ip−Iq)2 dist(p,q)·2σ2 ). Here Ip is the intensity of pixel p, and dist(p, q) is the Euclidean distance between p and q. Obse... |

239 | Learning to detect a salient object.
- Liu, Sun, et al.
- 2007
(Show Context)
Citation Context ...ntation will be made available on our web site. 5 Application to Salient Object Segmentation To show that regular superpixels are useful, we evaluated them for salient object segmentation, similar to =-=[28]-=-. The goal is to learn to segment a salient object(s) in an image. We use Berkeley dataset [15], 200 images for training and 100 for testing. Using human marked boundaries as a guide, we manually sele... |

215 | Recovering human body configurations: Combining segmentation and recognition
- Mori, Ren, et al.
- 2004
(Show Context)
Citation Context ...f l is the “orange” label, then S(l) is the set of pixels covered by the orange square. Label l can be assigned only to pixels in S(l). Therefore the data term is: Dp(l) = { 1 if p ∈ S(l) ∞ otherwise =-=(2)-=- We have to decide how many patches to use and how to spread them out in the image. We address these issues after the energy function is completely specified. We now discuss the smoothness term. To be... |

154 |
Efficient Approximate Energy Minimization via Graph Cuts.
- Boykov, Veksler, et al.
- 2001
(Show Context)
Citation Context ...y to include explicitly into [14]. Another advantage is optimization. Turbopixels are based on level set evolution [17], which is known to have numerical stability issues. We optimize with graph cuts =-=[18]-=-, which is known to perform well [19]. Our running time is better. Last, but not least, our approach naturally transfers to 3D for “supervoxel” segmentation of video. An interesting work on superpixel... |

150 |
Sethian J.A.: Fronts propagating with curvature-dependent speed: Algorithms Based on Hamilton-Jacobi Formulations.
- Osher
- 1988
(Show Context)
Citation Context ... that encourages intensity homogeneity inside a superpixel, not something that is easy to include explicitly into [14]. Another advantage is optimization. Turbopixels are based on level set evolution =-=[17]-=-, which is known to have numerical stability issues. We optimize with graph cuts [18], which is known to perform well [19]. Our running time is better. Last, but not least, our approach naturally tran... |

116 | Improving spatial support for objects via multiple segmentations
- Malisiewicz, Efros
- 2007
(Show Context)
Citation Context ...at we would have gotten performance similar to ours using turbopixels [14] or superpixels from [1], but our computational time is much better. Our results in this section are consistent with those of =-=[6]-=-, who show that having more accurate spatial support (more accurate superpixels) improves object segmentation. We also investigate whether the results from classification can be further improved by sp... |

107 | VENKATESH S.: Video abstraction: A systematic review and classification.
- TRUONG
- 2007
(Show Context)
Citation Context ...oxels can be useful, potentially, for medical image and for video processing. In particular, for video processing, there is an interest in coherent 3D segmentation for video abstraction and animation =-=[25, 26]-=-. First we create a 3D volume by stacking the frames together, Fig. 5, left. Analogously to the 2D case, we cover the 3D volume by overlapping 3D blocks. For clarity, in Fig. 5 we show only a few non-... |

99 | TurboPixels: Fast Superpixels Using Geometric Flows.
- Levinshtein, Stere, et al.
- 2009
(Show Context)
Citation Context ... explicit constraints on length. A large superpixel with a highly irregular shape is likely to straddle more than one object. Fig. 1. From left to right: meanshift [11], graph based [12], turbopixels =-=[14]-=-, NC superpixels [1]. Implementation was obtained from the authors’ web sites. There are advantages to superpixels with regular shapes and sizes, such as those in Fig. 1, on the right. A regular shape... |

83 | Guiding Model Search using Segmentation,”
- Mori
- 2005
(Show Context)
Citation Context ...controlled. The normalized cuts algorithm [13] can be adapted to compute superpixels that are regular in size and shape [1], see Fig. 1. Many methods that need regular superpixels use normalized cuts =-=[9, 1, 2, 4, 5]-=-. However, NC superpixels [1] are very expensive, and have the following unappealing property, noticed by [14]. The smaller is the size of target superpixels, the longer the computation takes. Our wor... |

73 | Learning and Incorporating Top-Down Cues in Image Segmentation.
- He, Zemel, et al.
- 2006
(Show Context)
Citation Context ...controlled. The normalized cuts algorithm [13] can be adapted to compute superpixels that are regular in size and shape [1], see Fig. 1. Many methods that need regular superpixels use normalized cuts =-=[9, 1, 2, 4, 5]-=-. However, NC superpixels [1] are very expensive, and have the following unappealing property, noticed by [14]. The smaller is the size of target superpixels, the longer the computation takes. Our wor... |

57 | VideoTrace: Rapid Interactive Scene Modelling from Video
- HENGEL, DICK, et al.
(Show Context)
Citation Context ...fficiency. If one needs to compute a property that stays approximately constant for an object, then superpixel representation is more efficient since the total number of primitives is greatly reduced =-=[9]-=-. Computational efficiency also comes from a reduction in the number of hypothesis. Instead of exhaustive examining of all rectangular patches [10], an alternative is2 Superpixels and Supervoxels in ... |

51 | Object recognition by integrating multiple image segmentations - Pantofaru, Schmid, et al. - 2008 |

15 |
Learning to find brightness and texture boundaries in natural images
- Martin, Fowlkes, et al.
(Show Context)
Citation Context ...els [14] and NC superpixels [1]. Visually the results are similar, except the NC superpixels appear to have smoother boundaries. This is because in [1] they use a sophisticated boundary detector from =-=[27]-=-. We could incorporate this too in our framework, but it is rather expensive, it takes approximately 30 seconds to compute boundaries for one image. In Fig. 7, we show some of our segmentations. The t... |

11 |
M.F.: Video tooning
- Wang, Xu, et al.
- 2004
(Show Context)
Citation Context ...oxels can be useful, potentially, for medical image and for video processing. In particular, for video processing, there is an interest in coherent 3D segmentation for video abstraction and animation =-=[25, 26]-=-. First we create a 3D volume by stacking the frames together, Fig. 5, left. Analogously to the 2D case, we cover the 3D volume by overlapping 3D blocks. For clarity, in Fig. 5 we show only a few non-... |

10 |
J.: Lattice cut - constructing superpixels using layer constraints
- MOORE, PRINCE, et al.
- 2010
(Show Context)
Citation Context ...is known to perform well [19]. Our running time is better. Last, but not least, our approach naturally transfers to 3D for “supervoxel” segmentation of video. An interesting work on superpixels is in =-=[20, 21]-=-. Their goal is somewhat different from ours. They seek superpixels that conform to a grid, which has storage and efficiency advantages. The work in [20] is based on greedy optimization, and [21] uses... |

2 | Class segmentation and objectlocalization with superpixel neighborhoods - Fulkerson, Vedaldi, et al. - 2009 |