IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE,

VOL. 25, NO. 6,

JUNE 2003

769

Short Papers

___________________________________________________________________________________________________

Expandable Bayesian Networks for 3D Object Description from Multiple Views and Multiple Mode Inputs
ZuWhan Kim, Member, IEEE Computer Society, and Ramakant Nevatia, Fellow, IEEE
AbstractComputing 3D object descriptions from images is an important goal of computer vision. A key problem here is the evaluation of a hypothesis based on evidence that is uncertain. There have been few efforts on applying formal reasoning methods to this problem. In multiview and multimode object description problems, reasoning is required on evidence features extracted from multiple images and nonintensity data. One challenge here is that the number of the evidence features varies at runtime because the number of images being used is not fixed and some modalities may not always be available. We introduce an augmented Bayesian network, the expandable Bayesian network (EBN), which instantiates its structure at runtime according to the structure of input. We introduce the use of hidden variables to handle correlation of evidence features across images. We show an application of an EBN to a multiview building description system. Experimental results show that the proposed method gives significant and consistent performance improvement to others. Index TermsMultiview object description, learning, uncertain reasoning, building description, Bayesian network.

modalities is that the number of evidence features that are used for reasoning varies at runtime because the number of images used is not fixed and some sensor data may not always be available. Therefore, a reasoning tool for varying numbers of evidence features and missing data is needed. We introduce an augmented Bayesian network, an expandable Bayesian network (EBN), which instantiates its structures at runtime according to the structure of the input. We first give an introduction to Bayesian networks in Section 2. Then, we formulate the problem and present expandable Bayesian networks in Section 3. Section 3 also contains a discussion of comparing EBNs with dynamic Bayesian networks (DBN) [7]. In Section 4, we show application of an EBN to a multiview and multimode 3D object description system called MVS (MultiView System for building detection and description [21]). We show experimental results in Section 5 and present the conclusion in Section 6.

BACKGROUND

1 INTRODUCTION
COMPUTING 3D object level descriptions from images is a key goal
of computer vision. The hypothesize and verify paradigm is a popular approach to solving this problem. Hypotheses of a higher level are constructed by grouping lower-level features; multiple hypotheses that explain the same part of an image or scene are possible. A choice among possibly competing and conflicting hypotheses is made by collecting supporting (or contradicting) evidence. A common problem in following this approach is that the decision to accept or reject a hypothesis must be based on evidence that is diverse in nature and uncertain. To make good decisions, two factors need to be considered: quality of the evidence and performance of the classifier. In this paper, we focus on the classifier. It has been a common practice to use ad hoc classification methods which lack generality; therefore, many computer vision systems suffer from severe overfitting where the system works well only on the small training data set (the demo examples) for which the parameters were manually tuned. There has been some recent work on use of uncertain reasoning and machine learning methods to 3D object description problems [25], [20], [18], but it is limited to the use of a single image. It is common to use more than one image of an object to retrieve 3D information efficiently. Often, nonintensity data, such as range data and multispectral imagery, are used as important cues to detect 3D structures. One challenge of using multiple images and

. Z. Kim is with the University of California at Berkeley, 1357 S. 46th St., Richmond, CA 94804. E-mail: zuwhan@cs.berkeley.edu. . R. Nevatia is with the Institute for Robotics and Intelligent Systems, University of Southern California, PHE 204, MC-0273, Los Angeles, CA 90089-0273. E-mail: nevatia@usc.edu. Manuscript received 2 Apr. 2002; accepted 7 Sept. 2002. Recommended for acceptance by R. Beveridge. For information on obtaining reprints of this article, please send e-mail to: tpami@computer.org, and reference IEEECS Log Number 116218.
0162-8828/03/$17.00 2003 IEEE Published by the IEEE Computer Society

We use capital letters, such as X and U , to denote random variables and lowercase letters, x and u, to denote certain assignments taken by those variables. Sets of variables are denoted by boldface capital letters, X , and assignments of theses sets are denoted by boldface lowercase letters, x. The statement P xju is used as a shorthand for P X xjU u. A Bayesian network [22] is a directed acyclic graph with random variables as its nodes. Edges represent causal relationships where all the child nodes are conditionally independent given a parent node. Each node has a conditional probability distribution (CPD) as a function of its parent node(s). The process of computing probabilities of some nodes given assignments of other nodes is the process of Bayesian inference. The computational complexity for an inference (of the worst-case query) in unconstrained network is known to be NP-hard with respect to the number of nodes [6]. A linear time algorithm for singly connected networks (or polytrees), in which there is at most one undirected path between any two nodes, is presented in [22] and [24]. When there is a multiple path or a loop, a clustering algorithm [22], [16] is often used. The structure of a Bayesian network can be determined by the known causal relations between nodes (causal Bayesian network) or learned by statistical methods [27]. When the nodes are discrete, CPDs can be estimated by simply collecting the statistics from examples. Continuous distributions are either represented by parametrized curves, such as Gaussian mixtures [11], [26], or discretized. The discretization method is easy to implement, does not require iterative search to find parameters (for observable nodes), and shows good performance [10]. In [12], Friedman and Goldszmidt suggested an approach to find good discretization. Hidden nodes (unobservable nodes) are introduced for several reasons. They can be used for compactness of the structure, to make causal dependencies explicit, or to satisfy the constraints of conditional independence [2]. Since the CPDs of hidden nodes and their children cannot be observed directly, several approaches have been proposed to obtain them [2], [5]. In [5], an EM (Expectation-Maximization) algorithm [8] is applied to find parameters for hidden nodes. In EM parameters of P XjU of node X and its parent(s) U are estimated by the following procedure (discrete case). In the expectation step, we compute

770

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE,

VOL. 25,

NO. 6,

JUNE 2003

Fig. 1. (a) A simple EBN with two nodes. Y is a repeatable node. (b) An instance for two images and (c) an instance for three images. P Y jX P Y1 jX P Y2 jX P Y3 jX (the commutativity constraint).

E Nx;u u

n X i

Fig. 2. (a) A use of a nonrepeatable hidden node for correlated evidence across

P x; ujdi ;

images (Y ). W and Z is correlated within an image. (b) Its instance for two views.

where Nx;u u is the number of cases where X x and U u in the learning examples d d 1 ; . . . ; dn . In the maximization step, we find P xju that maximizes P d (maximum-likelihood estimation); E Nx:u u P xju X ; E Nx0 ;u
x0 0

where x represents all the possible assignments of X . It is only until recent that Bayesian networks started to get wide attention in the computer vision community. Some of the early applications on object detection, description, and recognition problems can be found in [23], [25], [3], [20].

EXPANDABLE BAYESIAN NETWORK (EBN)

One of the difficulties of applying classifiers in multiview 3D object description problems is that the number of images being used is determined at runtime. Thus, the number of evidence features is not determined when the classifier is developed and learned. In this section, we introduce an expandable Bayesian network (EBN) to handle this. In this paper, we only consider two types of evidence features: ones directly from intensity images where the numbers of them vary according to the number of images (imagebased evidence features) and a fixed number (but can be missing) of evidence features obtained from nonintensity images (nonintensity evidence features). As the number of images is not determined, neither is their order. Therefore, the classifier should be commutative with respect to the order of the image-based evidence features (i.e., changing the order of images should not change the result). One approach to address this problem is to separately devise a classifier for image-based evidence and nonintensity evidence, and combine the results later. The same classifier should be used for each set of intensity-based evidence features to satisfy the commutativity requirement. For example, with four images, the same classifier is applied for all four sets of image-based evidence features (we get four classification results), another classifier is used for nonintensity evidence (one classification result), and five results are combined to get the final classification result. For the final combination, information fusion techniques such as majority voting, linear combinations (weighted sum), and certainty factor analysis [4] can be used. Majority voting (or Bagging [1]) gives reliable classification results when the number of the input classifiers is large. However, in the 3D object description problems, the number of the input classifiers (equal or smaller than the number of images + 1) can be as small as 2, and making a decision on their binary classification results may not be reliable. Linear combination can give more reliable results because it can use the likelihood values (not just binary results) of the input classification results. But, the choice of weights between the results from the image-based evidence and the one from the nonintensity

evidence is not clear. Certainty factor analysis [4] may be a good choice because it is commutative, combines any numbers of evidence, handles missing data, and deals with uncertainty (although it is rather ad hoc). The combination approach has two intrinsic limitations. First, the optimal combination among the results of the classifiers is not clear. Since the subclassifiers are learned separately, the final result may not be optimal even if the results of subclassifiers are. Another problem is that the correlation of evidence features across images is not considered. It is common that some of the evidence features are strongly correlated across images (an example is shown in Section 4). However, this correlation is not properly handled in the combination approach because it first combines the evidence within an image, then combines the results among them later. As an alternative, we propose the concept of an expandable classifier. In such a classifier, a rule (classifier template) is created when the system is developed. Then, at runtime, an actual classifier (instance) is generated according to the structure of given input data by the predefined template. A Bayesian network is suitable for augmentation as an expandable classifier because it consists of probabilistic relationships and gives an optimal decision given these relationships. We call this augmentation an expandable Bayesian network (EBN). An EBN has a set of repeatable nodes for the image-based evidence features. A repeatable node is instantiated into multiple copies according to the number of images. Fig. 1 shows an example EBN. Fig. 1a shows a network template, and Fig. 1b and Fig. 1c show example network instances for various numbers of images. For the nonintensity evidence features, simply nonrepeatable nodes are used since Bayesian networks naturally handle missing data. Our implementation is simple: We make an instance by copying repeatable nodes according to the structures of input. For each instance, the conditional probability distribution (CPD) and the edges (links) of the nodes are copied. When an edge of a repeatable node, say X , is connected to another repeatable node, Y , an instance of the edge from the ith node instance X i should be linked to the instance Y i of the same ith node instance. EBNs can handle correlated evidence across images. Fig. 2 shows the use of a nonrepeatable hidden node when an evidence feature, Y , is correlated across images. For the purpose of comparison, we also show W and Z which are correlated within an image. As a result of instantiation (Fig. 2b), for Y , the correlation is handled across images while, for W and Z , it is handled within an image. Instead of using hidden nodes to resolve correlation among evidence, one could add edges between the correlated nodes (such as in Dynamic Bayesian networks), but this technique is not suitable for 3D object description tasks networks because it does not preserve commutativity. More comparison with Dynamic Bayesian networks is discussed in the end of this section.

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE,

VOL. 25, NO. 6,

JUNE 2003

771

Fig. 3. (a) Aerial images taken from Fort Hood, Texas, (b) 3D building hypotheses generated, and (c) a desired classification result (positive hypotheses).

In Section 2, we discussed the learning of Bayesian network parameters (conditional probability distributions). Learning parameters of an EBN is similar. We can first apply learning to each instance (not directly to the template network), and determine the parameters of the template network by projecting the learned parameters of instances to a constraint space. Note that we have a constraint that wj wk , where wj and wk are the conditional probabilities of jth and kth instances of a repeatable node. We need to project the parameters of each and every instance into this ~0 w1 ; . . . ; wm , where m is the constraint space. Suppose w number of instances, is projected onto the constraint space and ~ w; . . . ; w. We get the projected parameter is w ~0 w ~ w ~ 0: w 3

also been used to combine evidence from a varying number of images. The difference comes from the use of nonrepeatable nodes. DBNs have been used for temporal reasoning, and causal links between temporally consecutive instances combine the information from the instances. On the contrary, EBNs use nonrepeatable (hidden or observable) nodes to combine information from all the instances at once (i.e., EBNs do not enforce a Markov assumption). EBNs also provide commutative inference on the instances.

AN EBN FOR A BUILDING DESCRIPTION SYSTEM

By applying simple algebra, we find that the projection corresponds to averaging all the conditional probabilities of the instances: W
m 1X wi : m i

Hidden nodes play an important role in an EBN. When we apply an EM algorithm to learn parameters of hidden nodes, we need to apply the projection, (4), to the result of (2). Unfortunately, (2) cannot be applied directly to an EBN because the structures of the instances of w are different for each input data d i . As an approximation, we reverse the order; we apply (2) to the result of (4). Then, we get E Nx;u u
n X m X P xj ; uj jd i i j

where xj and u j are the jth instances of a repeatable node X and its parent node(s) U . The instantiation and learning procedures of EBNs are different from those of dynamic Bayesian networks (DBNs), [7], which have

We present an application of an expandable Bayesian network for the task of automatic detection and construction of 3D models of buildings in aerial images. It is a task of interest for many applications such as for radiowave reachability tests for wireless communications, computer graphics, virtual reality, and mission planning. It has proven to be difficult to automate and has been an active research area [13]. We apply an EBN to a 3D building description system called MVS (Multi-View System for building detection and description [21]). MVS uses multiple images and sensor data to detect and describe 3D models of rectilinear buildings. Fig. 3a shows an example on image windows of the Fort Hood area. Line features are extracted from images and grouped into approximate rectangles to generate 3D building component hypotheses (Fig. 3b). These hypotheses are then verified using various evidence features (hypothesis verification). Fig. 3c shows a desired classification result for Fig. 3b. Finally, overlap analysis is applied on the positively classified hypotheses to get the final building components. We apply an EBN to the hypothesis verification stage. We collect evidence support both from intensity images and nonintensity data. Table 1 summarizes the evidence features used for the classification. Evidence extracted directly from intensity images consists of three parts, which are the evidence supports for the roof, the shadow, and the wall of a building. The roof

772

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE,

VOL. 25,

NO. 6,

JUNE 2003

TABLE 1 Evidence Features Used in MVS

Fig. 4. An expandable Bayesian network for MVS constructed based on manual correlation analysis.

evidence consists of evidence for the line supports, crossing lines, and standard deviation of the roof pixel intensities (to check the uniformness of the intensity). The shadow evidence consists of existence of strong/weak junctions, horizontal/vertical line supports, and intensity statistics. The wall evidence consists of evidence for vertical line supports and the baseline supports. For further details on the image-based evidence, see [21]. The evidence from nonintensity data consists of the coverages on the processed cues of HYDICE (HYper-spectral Digital Imagery Collection Experiment [15]) and DEM (Digital Elevation Map) were used. HYDICE gives very accurate cues, but is not an easily available source. See [14] and [15] for the details on the nonintensity evidence. Each and every evidence support was vectorized into continuous, binary, or ternary numbers for the classification. We know some of the semantic relationships among the evidence features as shown in Table 1. Therefore, we use a causal EBN, where the structure of the template network is determined by causal relationships. However, purely depending on our intuitive knowledge of causality may not be reliable. Therefore, we started with a naive-Bayesian-like structure where all the evidence features are children of the hypothesis node. Then, we manually performed a correlation analysis between evidence of the same category. For example an edge is added between SS and SV which showed a strong correlation within an image. We also performed a correlation analysis of the same evidence among views. For example, RS is strongly correlated across images because the intensity statistics of the roof are all similar regardless of the viewing angles. Therefore, we added a nonrepeatable hidden node to handle this correlation. We use an additional information Size (the size of a projected hypothesis given a view) and the correlations between evidence features and Size were analyzed. Fig. 4 shows an expandable Bayesian network constructed based on this correlation analysis. For details on the correlation analysis, see [17].

We implemented the EBN of Fig. 4 and tested with the data set described above. Continuous evidence features were discretized into five different levels and hidden nodes were set to be threevalued.1 The EM algorithm (5) was used for learning the parameters of the hidden nodes. We compare the results of the proposed EBN to the combination approach introduced in Section 3: A naive Bayesian classifier was used for nonintensity data (HYDICE and DEM) and, for the image-based evidence, both a naive Bayesian classifier and a Bayesian network were compared. Naive Bayesian classifiers show good performances for many classification problems [19], [9]. For the naive Bayesian classifier, evidence features with continuous values were also discretized into five different levels. For the Bayesian network, the same structure as the proposed EBN (Fig. 4), without the nonrepeatable hidden nodes, HYDICE and DEM, was used. For the final combination of the results, the certainty factor analysis was applied (Section 3). We present comparisons of the three classifiers, the EBN, the combination of naive Bayesian classifiers (CNBC), and the combination of Bayesian networks (CBN), both with and without the evidence from nonintensity data. We also present an experiment with synthesized evidence to compensate for the lack of large amount of real evidence from nonintensity data.

5.1

Evaluation Method

EXPERIMENTAL RESULTS AND DISCUSSION

To evaluate the suggested approach, we created a learning data set from aerial images. First, building hypotheses were generated from sets of two or three image windows of the same areas. From the building hypotheses, a learning dataset was created by displaying one hypothesis at a time to a human and asking for a decision on whether it is a true positive or not. One thousand two hundred and three learning examples were collected from 13 different image combinations of nine different locations; 781 of them were determined to be positive and 422 to be negative hypotheses. Eight hundred and four examples were from two views and 399 were from three views.

For the evaluation, we show trade off curves (ROC curves) between missed detection rate and false alarm rate.2 For Bayesian classifiers, we can generate ROC curves by simply modifying the threshold. We apply stratified k-fold cross validations [19] for statistical verification. In a k-fold cross validation, the data set is randomly divided into k small datasets (folds). For each fold, the outputs of the classifier which is trained with the other folds are collected. The final performance is obtained by combining the results of all the folds. To apply a k-fold cross validation to get an ROC curve, we first collect the posterior probabilities of the test data from all the folds. Then, we obtain missed detection rates and false alarm rates with various threshold values. To get statistical variations, we apply cross validations for a number of times and calculate the means and the standard deviations (confidence intervals) of missed detection rates for given false alarm rates. The resulting ROC curves in the following sections have small confidence intervals (typically less than one percentage point), which illustrates the reliability of the proposed evaluation method.
1. The discretization levels were set empirically. 2. Typical ROC curves are presented with detection rates and false alarm rates. However, in this application, the missed detection rates are considered to be more important.

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE,

VOL. 25, NO. 6,

JUNE 2003

773

Fig. 5. (a) Experiment with image-based evidence only and (b) with nonintensity evidence (200 examples).

5.2

Testing Image-Based Evidence Only

We compare the performance among the EBN, CNBC, and CBN. Stratified five-fold cross validations were repeated 10 times and ROC curves were obtained with confidence intervals. Evidence from nonintensity data was not used. Fig. 5a shows the resulting ROC curves. All three classifiers show radically improved performance over the original ad hoc classifier of [21]. We also find that the EBN shows significantly and consistently better performance than the other classifiers. For example, keeping the false alarm rate fixed at 6 percent, the EBN shows a 4.0 percentage point increase over that of the CNBC and a 1.9 percentage point improvement over CBN, while the confidence interval (shown as vertical bars in Fig. 5a) of the EBN is only a 0.66 percentage point. Only at 2 percent false alarm rate are the differences among the classifiers within the confidence

intervals. We believe that this is due to the stronger representational power of Bayesian networks (i.e., the use of the node, Size), accounting of correlation between nodes (SS and SV), and our formalism for the varying numbers of views.

5.3

Testing with Evidence from Nonintensity Data

To show the homogeneous combination of different type sets in EBN, we present the result with evidence features from nonintensity data (HYDICE and DEM) incorporated. Unfortunately, out of 1,203 examples, only 136 of them were covered by both HYDICE and DEM and 64 were covered by HYDICE only. Note that HYDICE coverage (HYDICE) enables much more accurate elimination of false alarms than that of the other evidence features which, as a result, dominates the classification results.3 As a result, all three classifiers show very good performances with HYDICE, and it is hard to distinguish the performances among the classifiers (Fig. 5b). It is hard to find statistically significant improvement of the EBN (small improvement given 2, 3, and 4 percent of false alarm within the confidence intervals). We next present an experiment with synthetic data to compensate for lack of large amount of real data. For given 1,203 examples, HYDICE and DEM scores were generated synthetically such that their means and the standard deviation is similar to other image-based evidence. This result is shown in Fig. 6. We find that, with evidence from nonintensity data, the improvement is still significant for most of the false alarm rates, which shows that EBN combines evidence from different type sets appropriately. For example, given 6 percent of false alarm rate, the performance improvements in missed detection rate of the EBN over the naive Bayesian classifier and the Bayesian network are 4.8 and 3.1 percentage points, respectively, which is much bigger than the confidence interval (a 0.56 percentage point).

CONCLUSION AND FUTURE WORK

We have introduced the use of formal reasoning and learning in a multiview and multimode 3D object description task. Our method
Fig. 6. A result when synthetic evidence is included. Stratified five-fold cross validations were repeated for 10 times.

3. However, it does not solve the whole problem because the process of obtaining HYDICE cues is not fully automatic.

774

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, [22] [23]

VOL. 25,

NO. 6,

JUNE 2003

provides a near-optimal decision given the quality of evidence. We have presented a methodology for using Bayesian networks and demonstrated their use in a building modeling task. An augmented Bayesian network, an expandable Bayesian network (EBN), which handles the combination of varying numbers of diverse evidence sets has introduced. We believe that the suggested methodology is general enough to be applied for various object detection and description tasks as well as other stereo or multiview image processing problems, and can provide a formal way to improve the generality of these systems.

[24] [25]

[26] [27]

J. Pearl, Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. San Francisco: Morgan Kaufmann, 1988. R.D. Rimey and C.M. Brown, Control of Selective Perception Using Bayes Nets and Decision Theory, Intl J. Computer Vision, vol. 12, no. 2, pp. 173207, 1994. S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach. PrenticeHall, 1995. S. Sarkar and K.L. Boyer, Using Perceptual Inference Networks to Manage Vision Process, Computer Vision and Image Understanding, vol. 62, pp. 27-46, 1995. V. Tresp, M. Haft, and R. Hofmann, Mixture Approximations to Bayesian Networks, Proc. 15th Conf. Uncertainty in Artificial Intelligence, 1999. T. Van Allen and R. Greiner, Model Selection Criteria for Learning Belief Nets: an Empirical Comparison, Proc. 17th Intl Conf. Machine Learning, pp. 1047-1054, 2000.

ACKNOWLEDGMENTS
This research was supported by a MURI subgrant from Purdue University under Army Research Office grant No. DAAH04-961-0444.
. For more information on this or any other computing topic, please visit our Digital Library at http://computer.org/publications/dlib.

REFERENCES
[1] [2] E. Bauer and R. Kohavi, An Empirical Comparison of Voting Classification Algorithms: Bagging, Boosting and Variants, Machine Learning, vol. 36, pp. 105-142, 1998. J. Binder, D. Koller, S. Russell, and K. Kanazawa, Adaptive Probabilistic Networks with Hidden Variables, Machine Learning, vol. 29, pp. 213-244, 1997. M. Brand, N. Oliver, and A. Pentland, Coupled Hidden Markov Models for Complex Action Recognition, Proc. IEEE Conf. Computer Vision and Pattern Recognition, pp. 995-999, 1997. Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project. B. Buchanan and E. Shortliffe eds., Addison Wesley, 1984. D.M. Chickering and D. Heckerman, Efficient Approximations for the Marginal Likelihood of Bayesian Networks with Hidden Variables, Proc. 12th Conf. Uncertainty in Artificial Intelligence, pp. 158-168, 1996. G.F. Cooper, The Computational Complexity of Probabilistic Inference Using Bayesian Belief Networks, Artificial Intelligence, vol. 42, pp. 393-405, 1990. T. Dean and K. Kanazawa, A Model for Reasoning about Persistence and Causation, Artificial Intelligence, vol. 93, pp. 1-27, 1989. A. Dempster, N. Laird, and D. Rubin, Maximum Likelihood from Incomplete Data via the EM Algorithm, J. Royal Statistical Soc. B, vol. 39, pp. 1-38, 1977. P. Domingos and M. Pazzani, On the Optimality of the Simple Bayesian Classifier Under Zero-One Loss, Machine Learning, vol. 29, pp. 103-130, 1997. J. Dougherty, R. Kohavi, and M. Sahami, Supervised and Unsupervised Discretization of Continuous Features, Proc. 12th Intl Conf. Machine Learning, 1995. R.O. Duda and P.E. Hart, Pattern Classification and Scene Analysis. John Wiley and Sons, 1973. N. Friedman and M. Goldszmidt, Discretizing Continuous Attributes while Learning Bayesian Networks, Proc. 13th Intl Conf. Machine Learning, pp. 157-165, 1996. Computer Vision and Image Understanding: Special Issue on Automatic Building Extraction from Aerial Images. A. Gruen and R. Nevatia, eds., vol. 72, no. 2, 1998. A. Huertas, Z. Kim, and R. Nevatia, Use of Cues from Range Data for Building Modeling, Proc. DARPA Image Understanding Workshop, pp. 577582, 1998. A. Huertas, R. Nevatia, and D. Landgrebe, Use of Hyperspectral Data with Intensity Images for Automatic Building Modeling, Proc. Second Intl Conf. Information Fusion, pp. 680-687, 1999. F.V. Jensen, S.L. Lauritzen, and K.G. Olesen, Bayesian Updating in Recursive Graphical Models by Local Computations, Technical Report R 89-15, Institute for Electronic Systems, Dept. of Math. and Computer Science, Univ. of Aalborg, Denmark, 1989. Z. Kim, Multi-View 3D Object Description with Uncertain Reasoning and Machine Learning, PhD Thesis, Computer Science Dept., Univ. of Southern California, 2001. Z. Kim and R. Nevatia, Uncertain Reasoning and Learning for Feature Grouping, Computer Vision and Image Understanding, vol. 76, pp. 278-288, 1999. R. Kohavi, A Study of Cross-Validation and Bootstrap for Accuracy Estimation and Model Selection, Proc. 14th Intl Joint Conf. Artificial Intelligence, pp. 1137-1143, 1995. M.A. Maloof, P. Langley, T.O. Binford, and R. Nevatia, Generalizing over Aspect and Location for Rooftop Detection, Proc. 14th IEEE Workshop Applications of Computer Vision, pp. 194-199, 1998. S. Noronha and R. Nevatia, Detection and Description of Buildings from Multiple Aerial Images, IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 23, no. 5, pp. 501-518, 2001.

[3]

[4]

[5]

[6]

[7] [8]

[9]

[10]

[11] [12]

[13]

[14]

[15]

[16]

[17]

[18]

[19]

[20]

[21]