## Salient local visual features for shape-based 3d model retrieval (2008)

Venue: | In SMI |

Citations: | 42 - 8 self |

### BibTeX

@INPROCEEDINGS{Ohbuchi08salientlocal,

author = {Ryutarou Ohbuchi and Kunio Osada and Takahiko Furuya and Tomohisa Banno},

title = {Salient local visual features for shape-based 3d model retrieval},

booktitle = {In SMI},

year = {2008}

}

### OpenURL

### Abstract

In this paper, we describe a shape-based 3D model retrieval method based on multi-scale local visual features. The features are extracted from 2D range images of the model viewed from uniformly sampled locations on a view sphere. The method is appearance-based, and accepts all the models that can be rendered as a range image. For each range image, a set of 2D multi-scale local visual features is computed by using the Scale Invariant Feature Transform [22] algorithm. To reduce cost of distance computation and feature storage, a set of local features describing a 3D model is integrated into a histogram using the Bag-Of-Features approach. Our experiments using two standard benchmarks, one for articulated shapes and the other for rigid shapes, showed that the methods achieved the performance comparable or superior to some of the most powerful 3D shape retrieval methods. KEYWORDS: Content-based retrieval, multi-scale feature, bag-offeatures,

### Citations

5629 | Distinctive image features from scale-invariant keypoints
- Lowe
- 2004
(Show Context)
Citation Context ...based, and accepts all the models that can be rendered as a range image. For each range image, a set of 2D multi-scale local visual features is computed by using the Scale Invariant Feature Transform =-=[22]-=- algorithm. To reduce cost of distance computation and feature storage, a set of local features describing a 3D model is integrated into a histogram using the Bag-OfFeatures approach. Our experiments ... |

2666 | Modern Information Retrieval
- Baeza-Yates, Ribeiro-Neto
- 1999
(Show Context)
Citation Context ... the class C k . R-precision is in fact the same as the First-Tier measure used by Shilane, et al [PSB2004]. We use the term Rprecision, which is commonly used in the information retrieval literature =-=[3]-=-. In the first set of experiments, we evaluated the influence the number of views and the number of vocabulary size have on the retrieval performance. In the second set of experiments, we used the bes... |

1023 | Video Google: a text retrieval approach to object matching in videos
- Sivic, Zisserman
- 2003
(Show Context)
Citation Context ...al geometric shape of the models. The BoF approach is one of the most popular and powerful methods to compute distance among sets, or bags of features in the field of object recognition for 2D images =-=[33, 6, 9, 43]-=-. The BoF approach is inspired by the bag-of-words approach in text retrieval. Typically, the approach encodes a given local feature into one of several hundreds to thousands of visual word, by using ... |

628 | Visual categorization with bags of keypoints - Csurka, Dance, et al. - 2004 |

388 | PCA-SIFT: a more distinctive representation for local image descriptors - Ke, Sukthankar - 2004 |

321 |
Topology matching for fully automatic similarity estimation of 3D shapes
- Hilaga, Shinagawa, et al.
- 2001
(Show Context)
Citation Context ...considered the same if the figure is standing, running, or crouching. A group of methods that aim at articulation invariance uses topological approach to 3D shape comparison. Methods by Hilaga, et al =-=[13]-=- and by Tung et al [36] used extended Reeb graph. Methods by Biasotti [4] and Siddiqi et al [32] used medial-axis or skeleton-like representation and graph-based matching. These methods, however, are ... |

247 | Learning object categories from google image search
- Fergus, Li, et al.
- 2005
(Show Context)
Citation Context ...al geometric shape of the models. The BoF approach is one of the most popular and powerful methods to compute distance among sets, or bags of features in the field of object recognition for 2D images =-=[33, 6, 9, 43]-=-. The BoF approach is inspired by the bag-of-words approach in text retrieval. Typically, the approach encodes a given local feature into one of several hundreds to thousands of visual word, by using ... |

217 | Rotation invariant spherical harmonic representation of 3d shape descriptors
- Kazhdan, Funkhouser, et al.
- 2003
(Show Context)
Citation Context ...e Princeton Shape Benchmark (PSB) [30]. We compared the BF-SIFT with six other methods, which are, the Individual Match SIFT (IM-SIFT) algorithm, the LFD [5], the Spherical Harmonics Descriptor (SHD) =-=[19]-=-, and our implementations of the D2 Shape Distribution [23], the Absolute Angle Distance histogram (AAD) [25], and the Surflet-Pair Relation Histograms (SPRH) [42]. Our experiments showed that the pro... |

215 | Categorization by learned universal visual dictionary
- Winn, Criminisi, et al.
- 2005
(Show Context)
Citation Context ...4 - 6, 2008. We have experimentally evaluated the proposed Bag-of-features SIFT (BF-SIFT) method. To evaluate retrieval performance for articulated models, we used the McGill 3D Shape Benchmark (MSB) =-=[43]-=-. To evaluate retrieval performance for rigid models, we used the Princeton Shape Benchmark (PSB) [30]. We compared the BF-SIFT with six other methods, which are, the Individual Match SIFT (IM-SIFT) a... |

124 |
On Bending Invariant Signatures for Surfaces
- Elad, Kimmel
(Show Context)
Citation Context ...raph-based matching. These methods, however, are complex and difficult to compute under the presence of geometrical or topological error and noise. Another group of methods, e.g., the methods by Elad =-=[7]-=-, Jain [16, 17], and Ran Gal [10, 11] uses curvature and other local geometrical and/or topological properties of manifold surfaces as the feature for pose invariant shape comparison. The method by Ja... |

116 | COHEN-OR D.: Salient geometric features for partial shape matching and similarity
- GAL
- 2006
(Show Context)
Citation Context ...ds, however, are complex and difficult to compute under the presence of geometrical or topological error and noise. Another group of methods, e.g., the methods by Elad [7], Jain [16, 17], and Ran Gal =-=[10, 11]-=- uses curvature and other local geometrical and/or topological properties of manifold surfaces as the feature for pose invariant shape comparison. The method by Jain et al [Jain06, Jain07] employs a j... |

82 | Tools for 3D-object retrieval: Karhunen–Loeve transform and spherical harmomics - Vranic, Saupe, et al. - 2001 |

73 | 3d Model Retrieval
- Vranic, Saupe
- 2000
(Show Context)
Citation Context ...l-PCA, is integrated by taking the minimum of the two. One of the pose normalization methods, called mass-PCA, is based on inertial moment of a mass uniformly distributed on the surfaces of the model =-=[41]-=-. We approximate the mass by quasiMonte-Carlo sampling the surfaces of the model with points having unit mass, the method we used for the AAD feature [25]. After the points are generated, a 3x3 covari... |

70 | Three-dimensional shape searching: state-of-the-art review and future trends. Computer-Aided Design
- Iyer, Jayanti, et al.
- 2005
(Show Context)
Citation Context ...r medical diagnostics, and for mechanical or architectural design. The need to organize these 3D models, for example for effective reuse, has prompted research into shape-based retrieval of 3D models =-=[35, 15]-=-. A 3D shape comparison method must satisfy several requirements for invariance. A typical set of requirements includes (1) invariance to similarity transformations, (2) invariance to shape representa... |

66 |
M.: Using spin-images for efficient multiple model recognition in cluttered 3-d scenes
- JOHNSON, HEBERT
- 1999
(Show Context)
Citation Context ... features associated with the relation of parts, for example, concavities in the shape. Yet another class of approaches uses a set of local features to achieve pose invariance for 3D model comparison =-=[18, 2, 14, 34, 21, 31]-=-. This class of method typically samples the surface of the model by using either 2D [18, 21, 2, 14] or 3D [31] local features. The 2D local feature based retrieval algorithms [2, 21] employ Spin Imag... |

64 | Poseoblivious shape signature
- Gal, Shamir, et al.
- 2007
(Show Context)
Citation Context ...ds, however, are complex and difficult to compute under the presence of geometrical or topological error and noise. Another group of methods, e.g., the methods by Elad [7], Jain [16, 17], and Ran Gal =-=[10, 11]-=- uses curvature and other local geometrical and/or topological properties of manifold surfaces as the feature for pose invariant shape comparison. The method by Jain et al [Jain06, Jain07] employs a j... |

47 | Shape-Similarity Search of 3D Models by using Enhanced Shape Functions
- Ohbuchi, Minamitani, et al.
- 2003
(Show Context)
Citation Context ...first renders asset of range images of the model from multiple view directions about the model, as in the Light Field Descriptor (LFD) [5] or the Multiple-Orientation Depth Fourier Descriptor (MODFD) =-=[25]-=-. To extract local features from each range image, the proposed method uses the Scale Invariant Feature Transform (SIFT) algorithm proposed by Lowe [Lowe04]. As each depth image yields a few dozen fea... |

47 | Surflet-Pair-Relation Histograms: A Statistical 3D-Shape Representation for Rapid Classification
- Wahl, Hillenbrand, et al.
(Show Context)
Citation Context ...Spherical Harmonics Descriptor (SHD) [19], and our implementations of the D2 Shape Distribution [23], the Absolute Angle Distance histogram (AAD) [25], and the Surflet-Pair Relation Histograms (SPRH) =-=[42]-=-. Our experiments showed that the proposed BF-SIFT performed the best among those compared in retrieving articulated 3D models of the MSB. The BF-SIFT produced R-precision=75% for the MSB, compared to... |

38 | Zhang H.: A spectral approach to shape-based retrieval of articulated 3D models
- Jain
- 2007
(Show Context)
Citation Context ... matching. These methods, however, are complex and difficult to compute under the presence of geometrical or topological error and noise. Another group of methods, e.g., the methods by Elad [7], Jain =-=[16, 17]-=-, and Ran Gal [10, 11] uses curvature and other local geometrical and/or topological properties of manifold surfaces as the feature for pose invariant shape comparison. The method by Jain et al [Jain0... |

35 | Parts-based 3d object classification
- Huber, Kapuria, et al.
(Show Context)
Citation Context ... features associated with the relation of parts, for example, concavities in the shape. Yet another class of approaches uses a set of local features to achieve pose invariance for 3D model comparison =-=[18, 2, 14, 34, 21, 31]-=-. This class of method typically samples the surface of the model by using either 2D [18, 21, 2, 14] or 3D [31] local features. The 2D local feature based retrieval algorithms [2, 21] employ Spin Imag... |

30 |
B.: Sub-part correspondence by structural descriptors of 3D shapes
- Biasotti, Marini, et al.
- 2006
(Show Context)
Citation Context ...oup of methods that aim at articulation invariance uses topological approach to 3D shape comparison. Methods by Hilaga, et al [13] and by Tung et al [36] used extended Reeb graph. Methods by Biasotti =-=[4]-=- and Siddiqi et al [32] used medial-axis or skeleton-like representation and graph-based matching. These methods, however, are complex and difficult to compute under the presence of geometrical or top... |

28 | Shape topics: a compact representation and new algorithms for 3D partial shape retrieval
- Liu, Zha, et al.
(Show Context)
Citation Context ... features associated with the relation of parts, for example, concavities in the shape. Yet another class of approaches uses a set of local features to achieve pose invariance for 3D model comparison =-=[18, 2, 14, 34, 21, 31]-=-. This class of method typically samples the surface of the model by using either 2D [18, 21, 2, 14] or 3D [31] local features. The 2D local feature based retrieval algorithms [2, 21] employ Spin Imag... |

27 | Distinctive regions of 3d surfaces
- SHILANE, FUNKHOUSER
- 2007
(Show Context)
Citation Context |

24 | Augmented reeb graphs for content-based retrieval of 3d mesh models. Shape Modeling Applications
- TUNG, SCHMITT
- 2004
(Show Context)
Citation Context ...the figure is standing, running, or crouching. A group of methods that aim at articulation invariance uses topological approach to 3D shape comparison. Methods by Hilaga, et al [13] and by Tung et al =-=[36]-=- used extended Reeb graph. Methods by Biasotti [4] and Siddiqi et al [32] used medial-axis or skeleton-like representation and graph-based matching. These methods, however, are complex and difficult t... |

16 |
On Visual Similarity Based 3D
- Chen, Tian, et al.
- 2003
(Show Context)
Citation Context ... set of local, multi-scale, salient, visual features. The method first renders asset of range images of the model from multiple view directions about the model, as in the Light Field Descriptor (LFD) =-=[5]-=- or the Multiple-Orientation Depth Fourier Descriptor (MODFD) [25]. To extract local features from each range image, the proposed method uses the Scale Invariant Feature Transform (SIFT) algorithm pro... |

16 |
R.C.: A Survey of Content based 3D
- Tangelder, Veltkamp
- 2008
(Show Context)
Citation Context ...r medical diagnostics, and for mechanical or architectural design. The need to organize these 3D models, for example for effective reuse, has prompted research into shape-based retrieval of 3D models =-=[35, 15]-=-. A 3D shape comparison method must satisfy several requirements for invariance. A typical set of requirements includes (1) invariance to similarity transformations, (2) invariance to shape representa... |

15 | Retrieving 3D shapes based on their appearance
- Ohbuchi, Nakazawa, et al.
- 2003
(Show Context)
Citation Context ...erformed and a set of sphericalharmonic features is extracted. As the spherical harmonic feature rotation invariant, the SHD achieves invariant to similarity transformation. The LFD [5] and the MODFD =-=[24]-=- uses a different method to achieve rotation invariance. Both methods perform normalization of position and scale. The rotation invariance is achieved by the combination of multiple-view rendering of ... |

10 | Mesh Retrieval by Components
- Tal, Zuckerberger
- 2004
(Show Context)
Citation Context ... the method accepts only closed mesh that has single connected component and has no internal structure. Another class of approach employs segmentation to partition a model into “meaningful” sub-parts =-=[34]-=-. The method then extracts a feature form each sub-part for a part-based, pose-invariant retrieval. Due to the decomposition, however, the method could miss features associated with the relation of pa... |

8 |
Learning semantic categories for 3D Model Retrieval
- Ohbuchi, Yamamoto, et al.
- 2007
(Show Context)
Citation Context ...the PSB. In retrieving rigid model of the PSB, some of the recent methods, such as the SPRH feature combined with the multiresolution feature extraction and Semi-Supervised Dimension Reduction (SSDR) =-=[27]-=- with R-precision=53%, would outperform the LFD, the SHD, and the BF-SIFT-42. While direct comparison has not been made, other recent methods such as the one by Napoléon, et al [39] or the one by Akgu... |

6 |
and Partha Niyogi, “Locality Preserving
- He
- 2003
(Show Context)
Citation Context ...ompared the performances of several linear dimension reduction algorithms, the PCA, Indipendent Compondent Analysis (ICA) using the FastICA algorithm [8] and the Locality Preserving Projections (LPP) =-=[12]-=-. In the set of parameters we experimented, the PCA performed the best. We use the subspace dimension of the PCA at which the contribution from the subspace is 99%. In pairing local features for indiv... |

5 |
A 2D Sketch-Based User Interface for 3D
- Pu, Lou, et al.
- 2005
(Show Context)
Citation Context ...etermines the direction of the model in the normalized coordinate system. Another pose normalization method is called normal-PCA. It again performs PCA but this time on the surface normal orientation =-=[29]-=-. Oftentimes, the mass-PCA performs better than the normalPCA. However, there are occasions in which normal-PCA does better. Our method thus computes two distances using the two pose normalization met... |

5 |
SIFT++ - A lightweight C++ implementation of SIFT”, http://vision.ucla.edu/~vedaldi/ code/siftpp/siftpp.html
- Vedaldi
(Show Context)
Citation Context ...entation of the local feature. We set the parameters to their defaults, which produces a 128D vector as a feature. To compute the SIFT features, we used the C++ implementation named SIFT++ by Vedaldi =-=[37]-=-. The SIFT algorithm typically produces anywhere from ten to a few hundreds of local features per image, depending on the image. According to our experiments, the numbers of features extracted average... |

4 | Press et al., Numerical Recipes in C-The Art - H - 1992 |

2 | Retrieving Articulated 3-D Models Using - Siddiqi, Zhang, et al. - 2008 |

2 |
et al., SHREC2006 3D Shape Retrieval Contest
- Veltkamp
(Show Context)
Citation Context ...ze the paper in Section 5. 2 RELATED WORK Recently, there is an increasing body of work on 3D model retrieval. Please refer to survey papers [35, 15] and reports of recent 3D model retrieval contests =-=[38, 39]-=- for comprehensive lists. There are many requirements for a shape comparison method. Most of the time, geometrical transformation invariance of the method to at least similarity transformation is expe... |

2 |
The McGill Shape Benchmark (2005). http://www.cim.mcgill.ca/shape/benchMark/ Qurey BF-SIFT with 42 views
- Zhang, Kaplow, et al.
(Show Context)
Citation Context ...he number of features. 4 EXPERIMENTS AND RESULTS We experimentally evaluated the retrieval performance of our approach by using two benchmark databases: the McGill university benchmark database (MSB) =-=[44]-=- for articulated shapes and Princeton Shape Benchmark (PSB) [30] for a set of diverse, rigid shapes. The MSB consists of 255 models in 10 classes. The MSB include such articulated shapes as “human”, “... |

1 |
Ohbuchi’s web
- unknown authors
(Show Context)
Citation Context ...,000 training features. All the other codebooks are generated by 50,000 training features. 80 70 60 of the D2, AAD, and the SPRH are available online as Windows XP (32bit) executables at our web site =-=[26]-=-. Figure 11a and Figure 11b shows, for the MSB and PSB, respectively the recall-precision plots for the five shape descriptors, the BF-SIFT, IM-SIFT, D2, SHD, and LFD descriptors For the articulated m... |