A survey of content based 3D shape retrieval methods
Summary (6 min read)
1. Introduction
- The advancement of modeling, digitizing and visualizing techniques for 3D shapes has led to an increasing amount of 3D models, both on the Internet and in domain-specific databases.
- Unlike text documents, 3D models are not easily retrieved.
- In contrast, content based 3D shape retrieval methods, that use shape properties of the 3D models to search for similar models, work better than text based methods [58].
- Matching and indexing are often part of the retrieval process.
2.1. 3D shape retrieval framework
- 1 consists of a database with an index structure created offline and an online query engine.
- Each 3D model has to be identified with a shape descriptor, providing a compact overall description of the shape.
- To efficiently search a large collection online, an indexing data structure and searching algorithm should be available.
- The similarity between two descriptors is quantified by a dissimilarity measure.
- Finally, the retrieved models can be visualized.
2.2. Shape representations
- An important issue is the type of shape representation(s) that a shape retrieval system accepts.
- Most of the 3D models found on the World Wide Web are meshes defined in a file format supporting visual appearance.
- In particular, they are represented by “polygon soups”, consisting of unorganized sets of polygons.
- Also, in general these models are not “watertight” meshes, i.e. they do not enclose a volume.
- By contrast, for volume models retrieval methods depending on a properly defined volume can be applied.
2.3. Measuring similarity
- In order to measure how similar two objects are, it is necessary to compute distances between pairs of descriptors using a dissimilarity measure.
- This property is very strong for a high-level shape descriptor, and is often not satisfied.
- This is not a severe drawback, if the loss of uniqueness depends on negligible details.
- When all the properties (i)-(iv) hold, the dissimilarity measure is called a metric.
- If a dissimilarity measure is a pseudo-metric, the triangle inequality can be applied to make retrieval more efficient [7, 83].
2.4. Efficiency
- For large shape collections, it is inefficient to sequentially match all objects in the database with the query object.
- Because retrieval should be fast, efficient indexing search structures are needed to support efficient retrieval.
- Since for query by example the shape descriptor is computed online, it is reasonable to require that the shape descriptor computation is fast enough for interactive querying.
2.5. Discriminative power
- A shape descriptor should capture properties that discriminate objects well.
- The judgement of the similarity of the shapes of two 3D objects is somewhat subjective, depending on the user preference or the application at hand.
- E.g. for solid modeling applications often topology properties such as the numbers of holes in a model are more important than minor differences in shapes.
- On the contrary, if a user searches for models looking visually similar the existence of a small hole in the model, may be of no importance to the user.
2.6. Partial matching
- In contrast to global shape matching, partial matching finds a shape of which a part is similar to a part of another shape.
- Partial matching can be applied if 3D shape models are not complete, e.g. for objects obtained by laser scanning from one or two directions only.
- Another application is the search for “3D scenes” containing an instance of the query object.
- Also, this feature can potentially give the user flexibility towards the matching problem, if parts of interest of an object can be selected or weighted by the user.
2.7. Robustness
- It is often desirable that a shape descriptor is insensitive to noise and small extra features, and robust against arbitrary topological degeneracies, e.g. if it is obtained by laser scanning.
- Also, if a model is given in multiple levels-ofdetail, representations of different levels should not differ significantly from the original model.
2.8. Pose normalization
- In the absence of prior knowledge, 3D models have arbitrary scale, orientation and position in the 3D space.
- Because not all dissimilarity measures are invariant under rotation and translation, it may be necessary to place the 3D models into a canonical coordinate system.
- For volume models it is natural to translate the center of mass to the origin.
- Eral not possible, because they have not to enclose a volume.
- Similar eigenvalues may imply an almost symmetrical mass distribution around an axis (e.g. nearly cylindrical shapes) or around the center of mass (e.g. nearly spherical shapes).
3. Shape matching methods
- In this section the authors discuss 3D shape matching methods.
- The authors divide shape matching methods in three broad categories: (1) feature based methods, (2) graph based methods and (3) other methods.
- Note, that the classes of these methods are not completely disjoined.
- A graph-based shape descriptor, in some way, describes also the global feature distribution.
- By this point of view the taxonomy should be a graph.
3.1. Feature based methods
- In the context of 3D shape matching, features denote geometric and topological properties of 3D shapes.
- Feature based methods can be divided into four categories according to the type of shape features used: (1) global features, (2) global feature distributions, (3) spatial maps, and (4) local features.
- Feature based Graph based Other Global features Spatial map Local features Global feature distribution Model graph Skeleton Reeb graph View based Volumetric error based Weighted point set based Deformation based Figure 3.
- Retrieving the k best matches for a 3D query model is equivalent to solving the k nearest neighbors problem.
- For this purpose, for each surface point a descriptor is used instead of a single descriptor.
3.1.1. Global feature based similarity
- Global features characterize the global shape of a 3D model.
- Zhang and Chen [88] describe methods to compute global features such as volume, area, statistical moments, and Fourier transform coefficients efficiently.
- Paquet et al. [67] apply bounding boxes, cords-based, moments-based and wavelets-based descriptors for 3D shape matching.
- Kazhdan et al. [42] describe a reflective symmetry descriptor as a 2D function associating a measure of reflective symmetry to every plane (specified by 2 parameters) through the model’s centroid.
- They improve the retrieval performance by an active learning phase in which a human annotator assigns attributes such as airplane, car, body, and so on to a number of sample models.
3.1.2. Global feature distribution based similarity
- The concept of global feature based similarity has been refined recently by comparing distributions of global features instead of the global features directly.
- Osada et al. [66] introduce and compare shape distributions, which measure properties based on distance, angle, area and volume measurements between random surface points.
- They evaluate the similarity between the objects using a pseudo-metric that measures distances between distributions.
- Since their method requires that a line segment can be classified as lying inside or outside the model it is required that the model defines a volume properly.
- Ohbuchi et al. [65] improved this method further by a multi-resolution approach computing a number of alpha-shapes at different scales, and computing for each alpha-shape their Absolute Angle-Distance descriptor.
3.1.3. Spatial map based similarity
- Spatial maps are representations that capture the spatial location of an object.
- The shape histograms are defined on concentric shells and sectors around a model’s centroid and compare shapes using a quadratic form distance measure to compare the histograms taking into account the distances between the shape histogram bins.
- Their method requires pose normalization to provide rotational invariance.
- Therefore, he compares his ray-based spherical harmonic method [85] and a variation of it using functions defined on concentric shells with the voxel-based spherical harmonics shape descriptor proposed by Funkhouser et al. [30].
- Their experimental results show that the eigenvalue method and the cover sequence method outperform the volume and solid-angle feature method.
3.1.4. Local feature based similarity
- Local feature based methods provide various approaches to take into account the surface shape in the neighbourhood of points on the boundary of the shape.
- Unfortunately, the method is limited to objects which contain no holes, i.e. have genus zero.
- Zaharia and Prêteux [87] describe the 3D Shape Spectrum Descriptor, which is defined as the histogram of shape index values, calculated over an entire mesh.
- They apply spin images to recognize models in a cluttered 3D scene.
- In the global matching stage, correspondences between similar sample points on the two shapes are found.
3.2. Graph based methods
- In general, the feature based methods discussed in the previous section take into account only the pure geometry of the shape.
- Graph based methods can be divided into three broad categories according to the type of graph used: (1) model graphs, (2) Reeb graphs, and (3) skeletons.
- Efficient computation of existing graph metrics for general graphs is not possible: computing the edit distance is NP-hard [90] and computing the maximal common subgraph [32] is even NP-complete.
- Polynomial solutions can be obtained for directed acyclic graphs such as shock graphs.
- It is obtained by exhaustively searching for the optimal deformation path between two 2D shapes, and using the cost of this path as a distance between two shapes.
3.2.1. Model graph based similarity
- Model graph based similarity methods are applicable to 3D solid models as produced by CAD most systems.
- The most dominant solid modeling representation methods are Boundary Representation (B-rep) and Constructive Solid Geometry (CSG).
- By contrast to the facets in meshes, the faces of a B-rep may be represented as freeform surfaces.
- McWerther et al. [53, 54, 55], and El-Mehalawi and Miller [26, 27] apply model signature graphs, that both model the topology of a shape model by a graph structure, and map a number of engineering features to a highdimensional feature vector.
- Therefore, McWerther et al. [53, 54, 55] apply approximate graph comparison using the spectrum of the graph.
3.2.2. Skeleton based similarity
- Sundar et al. [76] use as a shape descriptor a skeletal graph that encodes geometric and topological information.
- This topological signature vector is defined recursively over the subgraphs of the node using eigenvalues of their adjacency matrices.
- Sundar et al. [76] match two shapes by approximate comparison of their hierarchical skeletal graphs using Figure 5.
- They obtain a skeletal graph by a thinning algorithm iteratively eroding voxels until a one-voxel width skeleton is left.
- Their results show the feasibility of their approach for relatively small volume models.
3.2.3. Reeb graph based similarity
- Biasotti et al. [15] compare Reeb graphs obtained by using different quotient functions f and highlight how the choice of f determines the final matching result.
- Their method uses Reeb graphs based on a quotient function defined by an integral geodesic distance.
- Since for solid models topological insensitivity is important, they conclude that the Reeb graph technique requires some improvements.
- Bespalov et al. [13] present preliminary research on a modification of Hilaga’s method, which computes a scale-space decomposition of a shape, represented as a rooted undirected tree instead of a Reeb graph.
- In summary, Reeb graphs defined by a geodesic distance are suited for matching articulated objects, but they are sensitive to topological changes.
3.3.1. View based similarity
- The main idea of view based similarity methods is that two 3D models are similar, if they look similar from all viewing angles.
- Löffler [49] applies view based similarity to retrieve 3D models using a 2D query interface.
- The number of views of each object is kept small by clustering views, and by representing each cluster with one view, which is represented by a shock graph.
- Therefore, a lightfield descriptor is introduced, which compares ten silhouettes of the 3D shape obtained from ten viewing angles distributed evenly on the viewing sphere.
- The running time of the retrieval process is reduced by a clever multi-step approach supporting early rejection of non-relevant models.
3.3.2. Volumetric error based similarity
- Novotni and Klein [60] describe a geometry similarity approach to 3D shape matching based on calculating a volumetric error between one object and a sequence of offset hulls of the other object.
- A drawback of their method is that their dissimilarity measure is not symmetric and does not obey the triangle inequality.
- Sánchez-Cruz and Bribiesca present a method [69] relating the volumetric error between two voxelized shapes to a transportation distance measuring how many voxels have to move and how far to change one shape into another.
- Since in general these voxelized shapes will have many voxels, the computation of this transportation distance will be high.
3.3.3. Weighted point set based similarity
- Another approach is based on shape descriptors consisting of weighted 3D points.
- They match weighted point sets by a measure which does not obey the triangle inequality.
- Tangelder and Veltkamp [78] use as shape descriptor a weighted point sets consisting of points with a high curvature value.
- They compare weighted point sets using a variant of the Earth Mover’s distance, the proportional transportation distance, which obeys the triangle inequality [33].
- They utilize this multiresolution approximation to implement an algorithm to simultaneously align and compare two shapes.
3.3.4. Deformation based similarity
- A number of methods [20, 8] compare a pair of 2D shapes by measuring the amount of deformation required to register the shapes exactly.
- These methods depend on the natural arc length parameterization of their contours, which is not straightforwardly generalized to 3D.
- As a result, methods that apply deformation for shape recovery [79] or shape evolution [23] are very difficult to apply for 3D shape matching.
4. Overview and conclusions
- In this section the authors summarize their discussion on shape matching methods from the previous section and indicate directions for further research.
- Hence, these methods are less robust than feature based methods.
- The Princeton Shape Benchmark database contains 1,834 3D models downloaded from the web, subdivided into a training set and a test set, containing 907 models each, classified into 90 and 92 classes respectively.
- The vantage method [83] can be applied to compute an efficient index structure for pseudo-metrics that require much computing time, also known as Efficient indexing.
- Since the capabilities of feature based methods (fast computation, pseudo-metric, discriminative abilities, robustness) are orthogonal to the capabilities of graph based methods (partial matching, no normalization required), combining different approaches may produce more powerful shape matching methods.
Did you find this useful? Give us your feedback
Citations
894 citations
617 citations
613 citations
602 citations
573 citations
References
40,020 citations
34,239 citations
7,251 citations
6,693 citations
4,478 citations
Related Papers (5)
Frequently Asked Questions (16)
Q2. What are the future works in "A survey of content based 3d shape retrieval methods" ?
In this section the authors summarize their discussion on shape matching methods from the previous section and indicate directions for further research. Further research is needed to compare both approaches using the same benchmarks and the best PCA method. Because a feature vector is a point in a fixed d-dimensional space, two models can be compared fast by computing their distance in this space. Also, indexing is straightforward and retrieval can be implemented efficiently by nearest neighbour search.
Q3. What are the dominant solid modeling representation methods?
The most dominant solid modeling representation methods are Boundary Representation (B-rep) and Constructive Solid Geometry (CSG).
Q4. How do they compute a rotation invariant descriptor?
From the collection of spherical functions they compute a rotation invariant descriptor by (1) decomposing the function into its spherical harmonics, (2) summing the harmonics within each frequency, and computing the L2-norm for each frequency component.
Q5. Why is it necessary to place 3D models into a canonical coordinate system?
Because not all dissimilarity measures are invariant under rotation and translation, it may be necessary to place the 3D models into a canonical coordinate system.
Q6. What is the method of comparing shape histograms?
The shape histograms are defined on concentric shells and sectors around a model’s centroid and compare shapes using a quadratic form distance measure to comparethe histograms taking into account the distances between the shape histogram bins.
Q7. How do they improve the retrieval performance of 3D shapes?
They improve the retrieval performance by an active learning phase in which a human annotator assigns attributes such as airplane, car, body, and so on to a number of sample models.
Q8. How can the authors measure distances between two objects?
In order to measure how similar two objects are, it is necessary to compute distances between pairs of descriptors using a dissimilarity measure.
Q9. How many thumbnail images are used to obtain the descriptor of each 3D object?
In the preprocessing phase a descriptor of each 3D model is obtained by 13 thumbnail images of boundary contours of the 3D object as seen from 13 orthographic view directions.
Q10. What are the three broad categories of graph based methods?
Graph based methods can be divided into three broad categories according to the type of graph used: (1) model graphs, (2) Reeb graphs, and (3) skeletons.
Q11. What is the size of the topological signature vector?
Since a topological signature vector has a fixed size, the size of its shape descriptor is a constant multiplied by the number of nodes of the skeleton.
Q12. What are the disadvantages of local feature methods?
Compared to the methods from the first three categories, local feature methods, which compute feature value vectors for a number of surface points, matching is less efficient, efficient indexing is not straightforward, and the obtained dissimilarity measure is not obey the triangle inequality.
Q13. What is the generalization of the method used to compute a spherical harmonic?
note that their approach generalizes the method to compute voxel-based spherical harmonics shape descriptor, described by Funkhouser et al. [30], which is defined as a binary function on the voxel grid, where the value at each voxel is given by the negatively exponentiated Euclidean Distance Transform of the surface of a 3D model.
Q14. Why was it not possible to compare the results of different researchers?
Due to the lack of publicly available benchmarks in the past, it was not possible to compare the shape matching results obtained by different researchers.
Q15. What are the three approaches to provide a query object?
Three approaches can be distinguished to provide a query object: (1) browsing to select a new query object from the obtained results, (2) a direct query by providing a query descriptor, (3) query by example by providing an existing 3D model or by creating a 3D shape query from scratch using a 3D tool or sketching 2D projections of the 3D model.
Q16. How do they obtain a shape descriptor?
Dey et al. [24] present a method to obtain a descriptor of a shape, given by a point sample, by first decomposing the shape into its components.