## Using the Inner-Distance for Classification of Articulated Shapes (2005)

### Cached

### Download Links

Venue: | In Proc. CVPR |

Citations: | 57 - 9 self |

### BibTeX

@INPROCEEDINGS{Ling05usingthe,

author = {Haibin Ling and David W. Jacobs},

title = {Using the Inner-Distance for Classification of Articulated Shapes},

booktitle = {In Proc. CVPR},

year = {2005},

pages = {719--726}

}

### Years of Citing Articles

### OpenURL

### Abstract

We propose using the inner-distance between landmark points to build shape descriptors. The inner-distance is defined as the length of the shortest path between landmark points within the shape silhouette. We show that the innerdistance is articulation insensitive and more effective at capturing complex shapes with part structures than Euclidean distance. To demonstrate this idea, it is used to build a new shape descriptor based on shape contexts. After that, we design a dynamic programming based method for shape matching and comparison. We have tested our approach on a variety of shape databases including an articulated shape dataset, MPEG7 CE-Shape-1, Kimia silhouettes, a Swedish leaf database and a human motion silhouette dataset. In all the experiments, our method demonstrates effective performance compared with other algorithms. 1

### Citations

1347 | Shape matching and object recognition using shape contexts
- Belongie, Malik, et al.
- 2002
(Show Context)
Citation Context ...e quite different in (a) and (b), while almost the same in (b) and (c). The inner-distance is a natural replacement for the Euclidean distance in shape descriptors. We use it to extend shape contexts =-=[2]-=-. Based on the new descriptor, we design a dynamic programming method for silhouette matching that is fast and accurate. The proposed method is tested on a variety of shape databases. Excellent perfor... |

829 | Gradient-based learning applied to document recognition
- LeCun, Bottou, et al.
- 1998
(Show Context)
Citation Context ...s the average distance between a point on A and its most similar counterpart on B (in the sense of (9)). The SC+TPS is shown to be very effective for shape matching by tests [2] on the MNIST database =-=[11]-=-, MPEG7 CE-Shape-1, and others. 4.2 Extension of Shape Context To extend the shape context defined in (6), we redefine the bins with the inner-distance. The Euclidean distance is directly replaced by ... |

359 |
Object recognition by computer: The role of geometric constraints
- Grimson
- 1990
(Show Context)
Citation Context ...wo categories, supervised and unsupervised. The supervised methods explicitly build models for part structures through training. Then the models are used for retrieval tasks. Examples can be found in =-=[9, 6, 16]-=-. The unsupervised methods do not depend on explicit part models. For example, [1] showed that similarities of part structure can be captured without the explicit computation of part structure. [20, 1... |

249 | A new point matching algorithm for non-rigid registration
- Chui, Rangarajan
- 2003
(Show Context)
Citation Context ...ative. [13] extended the shape context by adding statistics of tangent vectors at landmark points. [22] suggested including a figural continuity constraint. [23] applied shape context and softassign =-=[4]-=- for fast and effective shape matching. In this paper, we extend the shape context by using the inner-distance to measure the spatial relations between points on shapes. Roughly speaking, current meth... |

162 | Recognizing objects in adversarial clutter: Breaking a visual captcha
- Mori, Malik
- 2003
(Show Context)
Citation Context ...elative spatial distribution (distance and orientation) of landmark points around feature points. Combined with the thin-plate-spline [3], the shape context is demonstrated to be very discriminative. =-=[13]-=- extended the shape context by adding statistics of tangent vectors at landmark points. [22] suggested including a figural continuity constraint. [23] applied shape context and softassign [4] for fas... |

159 | B.: Recognition of shapes by editing their shock graphs
- Sebastian, Klein, et al.
- 2004
(Show Context)
Citation Context ...6, 16]. The unsupervised methods do not depend on explicit part models. For example, [1] showed that similarities of part structure can be captured without the explicit computation of part structure. =-=[20, 18]-=- used shock graphs for shape comparison. Some other related works can be seen in [8]. 3 The Inner-Distance Now we describe the inner-distance. Consider two points x, y ∈ O, where O is a shape defined ... |

134 | Efficient and robust retrieval by shape content through curvature scale space
- Mokhtarian, Abbasi, et al.
- 1996
(Show Context)
Citation Context ...rent start points 6 Figure 9: Typical shape images from the MPEG7 CEShape-1, one image from each class. Table 2: Retrieval rate (bullseye) of different methods for the MPEG7 CE-Shape-1. Algorithm CSS =-=[12]-=- Visual Parts[10] SC+TPS[2] Score 75.44% 76.45% 76.51% Algorithm Curve Edit[17] Gen. Model[23] IDSC+DP Score 78.17% 80.03% 85.40% (ns = 8) are used in the DP matching and the penalty factor τ is set t... |

133 | Shape descriptors for non-rigid shapes with a single closed contour
- Latecki, Lakämper, et al.
- 2000
(Show Context)
Citation Context ...rates that our method is very effective for objects with articulated parts, while the shape context is not very suitable for this data set. 5.2 MPEG7 Shape Database The widely tested MPEG7 CE-Shape-1 =-=[10]-=- database consists of 1400 silhouette images from 70 classes. Each class has 20 different shapes (see Fig. 9 for some examples). The recognition rate is measured by the so called Bullseye test: For ev... |

119 | Flexible Syntactic Matching of Curves and Its Application to Automatic Hierarchical Classification of Silhouettes
- Gdalyahu, Weinshall
- 1999
(Show Context)
Citation Context ...es. 5.3 Kimia’s database The IDSC+DP is tested on two shape databases provided by Kimia’s group [19, 18]. The first database [19] contains 25 images from 5 categories (Fig. 11). It has been tested by =-=[2, 19, 7]-=-. In our experiment, 100 sample points are used for each silhouette, 5 distance bins and 12 orientation bins are used in IDSC, and ns = 4, τ = 0.3 are used in DP matching. The retrieval result is summ... |

83 | Symmetry-based indexing of image databases - Sharvit, Chan, et al. - 1998 |

27 |
Computer vision classification of leaves from Swedish trees
- Söderkvist
- 2001
(Show Context)
Citation Context ...tion of leaf shapes and texture make the problem very challenging. We use the Swedish leaf dataset from a leaf classification project at Linköping University and the Swedish Museum of Natural History =-=[21]-=-. The dataset contains isolated leaves from 15 different Swedish tree species, with 75 leaves per species. Fig. 13 shows some silhouette examples. Some initial classification work has been done in [21... |

19 |
Determining the similarity of deformable shapes. Vision Research
- Basri, Costa, et al.
- 1998
(Show Context)
Citation Context ...s for part structures through training. Then the models are used for retrieval tasks. Examples can be found in [9, 6, 16]. The unsupervised methods do not depend on explicit part models. For example, =-=[1]-=- showed that similarities of part structure can be captured without the explicit computation of part structure. [20, 18] used shock graphs for shape comparison. Some other related works can be seen in... |

18 |
Shape matching and recognition-using generative models and informative features
- Tu, Yuille
- 2004
(Show Context)
Citation Context ...ntext is demonstrated to be very discriminative. [13] extended the shape context by adding statistics of tangent vectors at landmark points. [22] suggested including a figural continuity constraint. =-=[23]-=- applied shape context and softassign [4] for fast and effective shape matching. In this paper, we extend the shape context by using the inner-distance to measure the spatial relations between points ... |

8 | Principal Warps: Thin-Plate-Splines and Decomposition of Deformations", PAMI - Bookstein - 1989 |

2 |
Object Detection Using the
- Schneiderman, Kanade
- 2002
(Show Context)
Citation Context ...wo categories, supervised and unsupervised. The supervised methods explicitly build models for part structures through training. Then the models are used for retrieval tasks. Examples can be found in =-=[9, 6, 16]-=-. The unsupervised methods do not depend on explicit part models. For example, [1] showed that similarities of part structure can be captured without the explicit computation of part structure. [20, 1... |

1 |
Matching and
- Petrakis, Diplaros, et al.
- 2002
(Show Context)
Citation Context ...ral to restrict the matching π with this order. To this end, we use dynamic programming to solve the matching problem. Dynamic programming is widely used for contour matching. Details can be found in =-=[1, 15]-=- for example. By default, the above method assumes the two contours are already aligned at their start and end points. Without this assumption, one simple solution is to try different alignments at al... |