A. R. Pope and D. G. Lowe. Learning object recognition models from images. In Proc. 4th Intl. Conf. Comp. Vis., pages 296--301. IEEE, 1993.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....Furthermore, none of these cost functions is invariant under scaling and or rotation of the point data. Other features such as key points and lines have been used to reduce the computational cost because a digital contour usually consists of much fewer features than of points. Pope and Lowe [109] modeled an object with a graph whose nodes represent the feature values and whose edges represent the spatial arrangement (symmetric, parallel) of the features. Objects are considered similar if their graphs are isomorphic; a similarity metric based on a probability density estimator is used to ....

....algorithm proposed in this thesis is a combination of the labeling and alignment methods. The algorithm achieves both efficient and accurate body part localization by making use of the advantages of the above two approaches and overcoming their disadvantages. The combined approach has been used in [109] for 2D 2D feature matching and 2D 3D matching in [85] The common feature is that the uncertainty information has been used for feature matching and for determining transformation uncertainty with which to predict the positions of adjacent features. The main differences of the RCR algorithm from ....

A. Pope and D. Lowe. Learning Object Recognition Models from Images. International Conference on Computer Vision, 296--301, 1993.

....position, curvature and orientation. This enables the de nition of relations such as collinearity, parallelism or symmetry (see gure 19) We see this as an important di erence to the above mentioned PCA based methods, neural network based systems (such as, e.g. 38] or Bayesian methods (such as [34, 27]) In this kind of systems the interpretation of lower and intermediate stages of representation becomes dicult. Histogram methods such as [26, 39, 42] can take advantage of the power of multiple 34 cues and have the ability of fast image processing and recognition. For instance [26] applies, in ....

....above, our symbolic representation of objects allows for application of relations such as collinearity and parallelism. Another example, which is essential for our learning algorithm, is the metric (4) There exist a variety of other systems making use of iconic representations (see, e.g. [5, 40, 29, 1, 34, 32]) In contrast to most of these icon based systems our icons (or symbols) have a parametrized description and symbolic meaning which allows for the de nition of such relations and also allows for the reconstruction of objects in an extremly sparse way. In the object recognition system [1] an ....

[Article contains additional citation context not shown here]

A.P. Pope and D.G. Lowe. Learning object recognition models from images. In T. Poggio and S. Nayar, editors, Early Visual Learning. 1995.

....reasonable to suppose that the data for the model construction problem are images representing different views of the object. In the literature, this problem has been approached for particular cases. For example, the construction of viewer centered 3 D models from intensity images is considered in [10]; GEON based (see Biederman s theory of recognition by components in [4] model construction has also been addressed, for example in [13] or in [14] In this paper, we present an approach for the synthesis of object centered models, considering partial models (views) as input data. Both the input ....

A.R. Pope and D.G. Lowe, `Learning Object Recognition Models from Images', in ICCV93, pp. 296--301, (1993).

....curvature and orientation. This enables the definition of relations such as collinearity, parallelism or symmetry (see figure 19) We see this as an important difference to the above mentioned PCA based methods, neural network based systems (such as, e.g. 38] or Bayesian methods (such as [34, 27]) In this kind of systems the interpretation of lower and intermediate stages of representation becomes difficult. Histogram methods such as [26, 39, 42] can take advantage of the power of multiple cues and have the ability of fast image processing and recognition. For instance [26] applies, in ....

....above, our symbolic representation of objects allows for application of relations such as collinearity and parallelism. Another example, which is essential for our learning algorithm, is the metric (4) There exist a variety of other systems making use of iconic representations (see, e.g. [5, 40, 29, 1, 34, 32]) In contrast to most of these icon based systems our icons (or symbols) have a parametrized description and symbolic meaning which allows for the definition of such relations and also allows for the reconstruction of objects in an extremly sparse way. In the object recognition system [1] an ....

[Article contains additional citation context not shown here]

A.P. Pope and D.G. Lowe. Learning object recognition models from images. In T. Poggio and S. Nayar, editors, Early Visual Learning. 1995.

....regions of an image, is the core process underlying a number of imaging tasks, including recognition ( is objectX in the image ) and identification ( which object is in the image ) as well as several forms of tracking ( find all moving objects of typeX in this sequence of images ) etc. PL95; HR96] Of course, it is critical that interpretation systems be accurate. It is typically important that the interpretation process also be fast: For example, to work in real time, an interpreter examining the frames of a motion picture will have only 1 24 of a second to produce an ....

A. Pope and D. Lowe. Learning object recognition models from images. In Early Visual Learning, 1995.

....of an object s relationship to the scene in which it is embedded. Traditional recognition techniques either rely on static feature measurements remaining stable in the presence of occlusions [19, 12, 3] or associate a likelihood of finding each feature based on off line appearance analysis [6, 26, 1, 11]. In most of these works, some occlusion is tolerated, but it is seldom dealt with explicitly. Instead, a match quality metric ranks potential matches, and matches with missing features are ranked lower. No provisions are made for explaining the absence of a feature in terms of interactions with ....

....by the junction of two faces on a 3D model. Under certain lighting conditions and viewing angles, that feature may appear prominent when imaged. Under other conditions it may not even be detectable. Based on this belief, methods for incorporating appearance information into a model have arisen [6, 26, 1, 11]. The appearance based approach derives additional information about stored models from a large set of training examples. In some cases, these images are artificially generated [6, 26] and in others they are real sensor images [1, 11] In each case, the goal is to extend the model representation ....

[Article contains additional citation context not shown here]

Arthur R. Pope and David G. Lowe. Learning Object Recognition Models from Images. In Tomaso Poggio and Shree Nayar, editor, Early Visual Learning. (Publisher not Known), 1995. http://www.cs.ubc.ca/spider/pope/home.html.

....is obtained by arranging the list of properties of an object as the components of a vector associated with the object. Similarity between objects is determined by the distance in feature space between the vectors associated with the objects. When local features are used (e.g. Pope and Lowe[41]) an object is represented by a graph with nodes representing the feature values and edges representing the spatial arrangement of the features. Objects are considered similar if their graphs are isomorphic. Methods that use features critically depend on the set of features extracted. A small ....

Pope, A. and D. Lowe, 1993, "Learning Object Recognition Models from Images," 4th Int. Conf. on Computer Vision:296--301.

....[26] Attribute value learning systems restrict the models to use global features, and their language is not expressive enough to easily describe the structure of objects. Thus, these approaches are not quite satisfactory. Among these systems, Connell and Brady [9] Segen [25] Pope and Lowe [20], and Zhang, et al. 31] have used graph oriented learning systems, whereas Cromwell and Kak [10] and Pellegreti et al. 18] have used the relational learning system INDUCE [15] These systems suffer from the limitations of INDUCE which can only learn rules in propositional form. In addition, ....

A. R. Pope and D. G. Lowe. Learning object recognition models from images. In Proceedings of the 4th Int. Conf. on Computer Vision, pages 296--301, 1993.

....can discrimi nate among the different classes of objects. The model is represented as a descriptor list of n ary relations among the features. This system relies heavily on good extractio n of object boundaries. The method is also difficult to apply to three dimensional objects. Pope and Lowe [PL93] have also designed a system that generates the models of objects represented by a graph. They construct a graph which represents the relations among the features derived by the perceptual organisation process. In a manner similar to Connell and Brady s, they try incrementally to generalise the ....

A. R. Pope and D. G. Lowe. Learning object recognition models from images. In Proceedings of the 4th Int. Conf. on Computer Vision, pages 296--301, 1993.

....to the thumb, to the index finger, etc. so we know which input units to clamp with which numbers. Success at the correspondence matching step is vital for correct classification. There has been much previous work on this topic (Connell and Brady 1987; Segen 1989; Huttenlocher and Ullman 1990; Pope and Lowe 1993) but a general solution has eluded the vision community. In this paper we propose a novel approach based on maximizing the probability of a set of models generating the given data. We show that neural networks trained to estimate the joint density between image features can be successfully used to ....

Pope, A. and D. Lowe (1993, May). Learning object recognition models from images.

....to find an object first find a small subset of features that predict the presence of the object. This general approach to recognition can be traced through many works including [34] 22] work on geometric hashing schemes [35] and on through a collection of excellent recent works [36] [37], 38] 39] 40] Grimson et al. provide a nice general analysis of the problem [41] The fundamental difficulty in designing indexing algorithms is efficiently finding reliable sets of domain independent indexing features. Hence, indexing is frequently solved using domain specific heuristics. ....

A.R. Pope and D.G. Lowe, "Learning object recognition models from images," in ICCV, 1993, pp. 296--301.

.... Many approaches to view based modeling represent each view as a collection of extracted features, such as extracted line segments, curves, corners, line groups, regions, or surfaces (Ikeuchi and Kanade [11] Burns and Kitchen [2] Ullman and Basri [30] Dickinson et al. 7] and Pope and Lowe [21]) The success of these view based recognition systems depends on the extent to which they can extract their requisite features. With real images of real objects in unconstrained environments, the extraction of such features can be both time consuming and unreliable. In contrast to the ....

....nature of their representations (although some offer limited invariance to particular transformations) However, the advantage of these approaches is that complex feature extraction, grouping, or abstraction is not required. Systems based on more invariant view based image descriptions, e.g. [11, 2, 30, 7, 21], have relied on complex feature extraction (e.g. edges, lines, regions, etc. which is not only unreliable but often requires domain specific parameter tuning. To address these shortcomings, we compute a scale space representation of an image in which image objects (homogeneous regions) are ....

A. Pope and D. Lowe. Learning object recognition models from images. In Proceedings, IEEE International Conference on Computer Vision, pages 296--301, Berlin, May 1993.

....with a wide variety of approaches. For example, work on comparing shapes with articulated parts are discussed in [7, 12, 13] Another approach to shape comparison describes shapes with a list of properties and their relations, using pattern matching techniques to determine similarity (e.g. see [27], and a review in [10] A third class of methods attempts to assign semantic interpretation to shapes, such as function, and make similarity judgements based on semantics (e.g. 37, 16, 29, 28] Also relevant to our work, 26] deform shapes by aligning the principal modes of their mass and ....

Pope, A. and D. Lowe, 1993, "Learning Object Recognition Models from Images," ICCV:296--301.

.... Most approaches to view based modeling represent each view as a collection of extracted features, such as extracted line segments, curves, corners, line groups, regions, or surfaces (Ikeuchi and Kanade [6] Burns and Kitchen [2] Ullman and Basri [16] Dickinson et al. 3] and Pope and Lowe [11]) The success of these view based recognition systems depends on the extent to which they can extract their requisite features. With real images of real objects in unconstrained environments, the extraction of such features can be not only time consuming but often unreliable. In contrast to the ....

A. Pope and D. Lowe. Learning object recognition models from images. In Proceedings, IEEE International Conference on Computer Vision, pages 296--301, Berlin, May 1993.

....IEEE Press: 781 785. 16] Milanese, R. Pun, T. Gil, S. Bost, J. M. 1994. Exploiting dynamic aspects of visual perception for objects recognition. In Proc. PerAc 94, From Perception to Action (Lausanne, Switzerland, September 7 9, 1994) IEEE Comp. Soc. Press, Los Alamitos, CA, 1994, 193 205. [17] Pope, A.R. and Lowe, D.G. 1993. Learning object recognition models from images. Proc. 4th Int. Conf. Comp. Vision ICCV 93, Berlin, Germany, May 11 13, 1993. 18] Pun, T. Bost, J. M. Milanese, R. Rauber, C. Startchik, S. 1994. Selecting relevant information and delaying irrelevant data for ....

....Complex tokens 4. 1 Learning of Relevant Primitives The practical usefulness of an object recognition system greatly depends on the use of learning techniques for building a knowledge base, that links the features extracted from the image to the cues needed for incremental recognition (e.g. 21] [17] [11] The machine learning problem is particularly difficult in the computer vision domain because there does not exist a well defined, predetermined set of attributes that compose the visual feature space. In the proposed approach, the keys for indexing a particular object in a data base are ....

Pope, A.R., and Lowe, D.G. 1993. Learning object recognition models from images. Proc. 4th Int. Conf. Comp. Vision ICCV'93, Berlin, Germany, May 11-13, 1993.

.... Many approaches to view based modeling represent each view as a collection of extracted features, such as extracted line segments, curves, corners, line groups, regions, or surfaces (Ikeuchi and Kanade [13] Burns and Kitchen [3] Ullman and Basri [33] Dickinson et al. 8] and Pope and Lowe [24]) The success of these view based recognition systems depends on the extent to which they can extract their requisite features. With real images of real objects in unconstrained environments, the extraction of such features can be both time consuming and unreliable. In contrast to the ....

....nature of their representations (although some offer limited invariance to particular transformations) However, the advantage of these approaches is that complex feature extraction, grouping, or abstraction is not required. Systems based on more invariant view based image descriptions, e.g. [13, 3, 33, 8, 24], have relied on complex feature extraction (e.g. edges, lines, regions, etc. which is not only unreliable but often requires domain specific parameter tuning. To address these shortcomings, we compute a scale space representation of an image in which image objects (homogeneous regions) are ....

A. Pope and D. Lowe. Learning object recognition models from images. In Proceedings, IEEE International Conference on Computer Vision, pages 296--301, Berlin, May 1993.

....objects. The techniques are typically difficult to extend to natural objects. Computer vision approaches to view based modeling fall broadly into two classes. First, there are feature based methods which represent each view as a collection of line segments, curves, corners, regions, etc. [22, 10, 11, 39]. The success of such methods depends largely on the extent to which the features are present and can be reliably extracted; once again they are not easily applied to natural objects. Second, a number of appearance based methods have emerged which essentially treat the raw image as a single ....

A. Pope and D. Lowe. Learning object recognition models from images. In Proceedings, IEEE International Conference on Computer Vision, pages 296--301, Berlin, May 1993.

....rare. Works have been done in related areas: In [29] for recognizing 3D objects from different viewpoints, a function mapping any viewpoint to a standard view is learned from a set of perspective views. In [40] a network structure is introduced for automated learning to recognize 3D objects. In [30], a numerical graph representation for an object model is learned from features computed from training images. A recent work [28] proposes a procedure of learning compatibility coefficients for relaxation labeling by minimizing a quadratic error function. Automated and optimal parameter estimation ....

Pope, A. R. and Lowe, D. G. (1993). "Learning object recognition models from images". In Proceedings of IEEE International Conference on Computer Vision, pages 296--301.

....deals with brightness images that are functions not only of shape but also other intrinsic scene properties such as reflectance and perpetually varying factors such as illumination. This observation has led to the exploration of view based approaches to object recognition (see [37] 43] 14] 44][38] for examples) It motivates us to take an extreme approach to visual representation. What we seek is not a representation of shape but rather appearance [19] encoded in which are brightness variations caused by three dimensional shape, surface reflectance properties, sensor parameters, and ....

A. R. Pope and D. G. Lowe, "Learning Object Recognition Models from Images, " Proc. of Fourth Int'l. Conf. on Computer Vision, pp. 296-301, Berlin, May 1993.

....used for recognition purposes. First, they were coded implicitly as assumptions about features at certain locations; later on, the explicit structural models were used [2] Statistical methods for knowledge acquisition and recognition of objects using non parametric estimates are described in [10]. The referred work relates learning of 3D objects to the automatic computation of an aspect graph of 3D objects, and does not explicitly model the 3D structure in the sense that the 3D coordinates of an object are included. The description of segmentation errors and feature deviations in ....

A.R. Pope and D.G. Lowe. Learning object recognition models from images. In Proc. 4. Int. Conf. on Computer Vision, pages 296--301, Berlin, May 1993. IEEE Press.

....variation. The learning procedure must generalize enough to overcome insignificant variation, but not so much as to confuse dissimilar objects. And the procedure used to identify modeled objects in images must tolerate the likely range of mismatch between model and image. In a previous paper [7], we considered the problem of learning a single characteristic view from a series of 2D training images. We are now extending that work by considering the problem of learning a set of characteristic views, automatically chosen to be sufficient for representing all aspects of a 3D object. Here we ....

....CV graph. The measure combines empirical statistics using the rules of Bayesian probability theory to judge the likelihood that a given pairing of CV tokens with image tokens corresponds to an actual instance of the object in the image. Since the graph similarity measure is described in detail in [7] we will only briefly summarize it here. It is essentially of the following form: g match between CV graph and image graph ( log L i matches j ( match between CV node i and image node j log L i is left unmatched ( unmatched CV node i The log likelihood of a match between an ....

[Article contains additional citation context not shown here]

A.R. Pope and D.G. Lowe, "Learning Object Recognition Models from Images," in: Proc. ICCV (1993) 296--301.

....test the approach. Section 7 discusses relevant work by others on this and similar problems, and section 8 summarizes the chapter s main ideas. Sections flagged by y contain technical details that can be safely skipped on a first reading. More information may be found in other recent publications [18, 19, 20]. 2 Representation Schemes 2.1 Image representation We represent an image in terms of discrete properties called features. Each feature has a particular type, a location within the image, and a vector of numeric attributes that further characterize it. A feature may, for example, be a segment ....

....B 20 56 6.7 Bottle B Bottle A 27 109 2.8 Bottle B Bottle B 27 56 20.7 Table 1: Results of matching each subimage of figure 7 with each bottle model. Features Paired is the proportion of model features paired. Match Quality is the value of the match quality measure, g(E; T ) Adapted from [19]. features. Some differences are due to shifts in the relative positions of features with changing viewpoint the seat and post remain fixed, for example, while the legs shift in various directions. Others are due to inherent differences in the accuracy of localizing various types of ....

[Article contains additional citation context not shown here]

A. R. Pope and D. G. Lowe. Learning object recognition models from images. In Proc. Int. Conf. Comput. Vision, pages 296--301, 1993.

....aligning transformation, is rated by the measure described in 3.3. One component of this measure estimates the probability that two features match given their respective position distributions and an aligning transformation; it is described in 3. 4; other components have been described previously [15]. The method of estimating a transformation from feature pairings is described in 3.6. A matching procedure, described in 3.5, uses the match quality measure and transformation estimator to match model features with image features. The matching procedure is used both to learn a model from training ....

....set of pairings between some model and image features, plus a transformation closely aligning paired features. We seek a match that maximizes both the number of features paired and the similarity of paired features. Our match quality measure quantifying these goals extends that reported in [15] to include an evaluation of how well the transformation aligns features. Pairings are represented by E = he 1 ; e 2 ; i, where e j = k if model feature j matches image feature k, and e j = if it matches nothing. H denotes the hypothesis that the modeled view of the object is present in the ....

[Article contains additional citation context not shown here]

A.R. Pope, D.G. Lowe. Learning object recognition models from images. In Proc. Int. Conf. Computer Vision, 296--301, 1993. 18

....of components from low level, generic primitives, through high level, object specific arrangements, to entire views of an object. Matching proceeds in stages from low level to high. Again, attributes are characterized by distributions to accommodate varation in appearance. Our previous work [11] used a graph matching method resembling both PREMIO and view description networks. Like PREMIO, we used a cost function and heuristic search to match graphs. Like Burns and Riseman, our graphs represented a range of features, from simple to complex, and each relation was characterized by a ....

....this measure. One component of it is an estimate of the probability that two features match, given their respective position distributions and an estimate of the viewpoint transformation; this component is described in section 3. 4 while other components have been described in previous work [11]. Section 3.5 describes how a viewpoint transformation is estimated from a set of feature pairings, and section Modeling Positional Uncertainty in Object Recognition 5 3.6 ties these pieces together in describing the procedure for finding a match. That matching procedure is used both for ....

[Article contains additional citation context not shown here]

A. R. Pope and D. G. Lowe, "Learning object recognition models from images," In Proc. ICCV (1993), pp. 296-301.

No context found.

A. R. Pope and D. G. Lowe. Learning object recognition models from images. In Proc. 4th Intl. Conf. Comp. Vis., pages 296--301. IEEE, 1993.

No context found.

A. Pope and D. Lowe. Learning object recognition models from images. In Proceedings, IEEE International Conference on Computer Vision, pages 296-- 301, Berlin, May 1993.

No context found.

A.P. Pope and D.G. Lowe. Learning object recognition models from images. In T. Poggio and S. Nayar, editors, Early Visual Learning. 1995.

No context found.

A. R. Pope and D. G. Lowe. Learning object recognition models from images. In Proc. 4th Intl. Conf. Comp. Vis., pages 296--301. IEEE, 1993.

No context found.

A. Pope and D. Lowe. Learning Object Recognition Models from Images. International Conference on Computer Vision, 296--301, 1993.

No context found.

A R Pope and D Lowe. Learning object recognition models from images. In Early Visual Learning, 1995.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC