J. R. Beveridge and E. M. Riseman. Optimal Geometric Model Matching Under Full 3D Perspective. Computer Vision and Image Understanding, 61(3):351 -- 364, 1995.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....parallelprocessing graphics hardware [8, 11, 13] Given the 3 D geometry of an object, the proposed recognition scheme autonomously determines the object s apparent size, image coordinates, and its pose in an uncalibrated image. When compared to previous research on geometry based recognition [3, 10, 9, 4], the presented scheme offers the advantages that no initial search parameter values must be provided, apparent object size may a priori be unknown, and that objects of arbitrary shape can be detected. In the succeeding section, the generation of multiple different object silhouette outlines is ....

J. Beveridge and E. Riseman. Optimal geometric modelmatching under full 3D perspective. Computer Vision, Graphics, and Image Processing (CVGIP): Image Understanding, 61(3):351--364, May 1995.

.... a proper model of its environment, an AMS can perform self localization as well as pose estimation with a single monocular video image by aligning model features like 3D points or 3D lines with the corresponding image features [1] A lot of work has been done to establish these correspondences [2, 3, 4] and thus to compute the world position of the camera or the relative pose of an object, respectively [5, 6, 7] The search for correspondences between model and image features has exponential complexity O(n m ) with n the number of possibly corresponding image features per model feature and m ....

J. R. Beveridge and E.M. Riseman. Optimal Geometric Model Matching Under Full 3d Perspective. In Second CADBased Vision Workshop, pages 54--63. IEEE Computer Society Press, 1994.

....minimized. 357 Some authors use the Gaussian error model. Sarachik and Grimson [14] analytically derived the probability of false positive and negative as a function of the number of model features, image features, and occlusions, under the assumption of 2D Gaussian noise. Beveridge and Riseman [15] presented a model based recognition system using 3D pose recovery during matching. Their algorithm is based on a random local search to find the globally optimal correspondence between image and model, with a high probability of success. A current set of correspondences is modified, using the ....

J.R. Beveridge and E.M. Riseman. Optimal geometric model matching under full 3d perspective. Computer Vision and Image Understanding, 61(3):351--364, 1995.

....on pose clustering in general include Stockman [11] who has studied it and compared it with some other paradigms, and Olson [10] who presents a method for efficient pose clustering. Related previous work for the 3D case of this application include Linnainmaa et al. 9] and Beveridge and Riseman [2]. 4.1 Controlling Complexity Since the number of triplet combinations will grow combinatorially with the amount of model and image features, we will need some means of either keeping these low and or selecting good candidate combinations. Here we will briefly discuss some possible methods and ....

J. Ross Beveridge and Edward M. Riseman, Optimal Geometric Model Matching under Full 3D Perspective, Computer Vision and Image Understanding, Vol 61, No 3, pp 351-364, May 1995

....in our case minimizing an error function #. In this way, the task is related to others we have considered before, such as optimal matching of 2D line segment models to cluttered and complex line data [4, 3] matching 3D line models to 2D image features assuming 3D perspective projection [1, 2, 8], optimal matching of 3D models to multi modal 5 data [11] and recent advances in combinatorial line matching using local search within genetic algorithms [14] Formally, let # be the set of # peaks extracted from the Semblance Velocity image by the procedure described above. The best solution ....

J. R. Beveridge and E. M. Riseman. Optimal Geometric Model Matching Under Full 3D Perspective. Computer Vision and Image Understanding, 61(3):351 -- 364, 1995.

....upon a single sensor. For example, models have been used for matching to 2D imagery [15, 9, 42] 3D range data [2, 4] as well as multi spectral imagery such as IR [34] and SAR [12] Typically these CAD systems rely on either 3D or 2D model geometry to constrain object location and appearance [28, 23, 21, 6, 10]. The more complex task of fusing data from sensors of different modalities, for instance range and optical, has also been addressed [30, 39, 43, 17, 24] However, this research area is still young. Fusing information from different modalities is complicated by many fac tors, including recovery ....

J. Ross Beveridge and Edward M. Riseman. Optimal Geometric Model Matching Under Full 3D Perspective. Computer Vision and Image Understanding, 61(3):351 -- 364, 1995. (short version in IEEE Second CAD-Based Vision Workshop).

....within an object recognition system. Hence it draws upon two literatures: object recognition for which there are many surveys [35, 16, 5] and CAD rendering [38, 18, 31] Most CAD based recognition systems focus upon a single sensor. For example, models have been used for matching to 2D imagery [15, 9, 42], 3D range data [2, 4] as well as multi spectral imagery such as IR [34] and SAR [12] Typically these CAD systems rely on either 3D or 2D model geometry to constrain object location and appearance [28, 23, 21, 6, 10] The more complex task of fusing data from sensors of different modalities, for ....

J. Ross Beveridge and Edward M. Riseman. Optimal Geometric Model Matching Under Full 3D Perspective. In Second CAD-Based Vision Workshop, pages 54 -- 63. IEEE Computer Society Press, February 1994. (Submitted to CVGIP-IU).

....models and multi sensor data. Both are demonstrated on real data. 1 Introduction Model based object recognition work has long emphasized the importance of aligning 3D object models to features extracted from sensed imagery [ BC82; Low91; GH91; HU90; HU88; BR92a; BR92b; Bev92a; Bev93; BR94; BHP94b; BR95 ] While model based approaches to Automatic Target Recognition have become much more common [ DVD93; GJSL90 ] direct incorporation of alignment into the recognition process is rare [ BJLP92 ] This paper presents algorithms and results from a project at Colorado State University which is ....

....] Huttenlocher [ HU90 ] demonstrated alignment based recognition under orthographic projection. Grimson s work [ Gri90 ] on constraint base matching has emphasized local feature compatibility to prune tree search and thus minimize tests of global alignment. Our past work [ BWR90; BWR91; Bev93; BR95 ] emphasized global alignment as a basis for match ranking and optimal matching to CCD sensor data. Local search through the match space constructs a globally consistent match by finding a sequence of successively better match hypotheses until one which is locally optimal is found. For a given ....

[Article contains additional citation context not shown here]

J. Ross Beveridge and Edward M. Riseman. Optimal Geometric Model Matching Under Full 3D Perspective. Computer Vision and Image Understanding, 61(3):351 -- 364, 1995. (short version in IEEE Second CAD-Based Vision Workshop).

....E fit drops between successive iterations falls below a preset threshold. Unsuccessful termination occurs if the total number of iterations exceeds a maximum number of iterations. The Levenberg Marquardt [ PFTV88 ] method has been found to be robust in our past single sensor pose work [ BR92b; BR94 ] and it is used here to find the optimal coregistration parameters. E fit ffi mo = w mo P no i=1 P 2 j=1 oi ii P e mij Theta N oi j Delta ffi mo N oi Delta Delta T mo j i P e mij Theta N oi j w mr P nr i=1 ri i M 2 moi ffi mo M moi ....

J. Ross Beveridge and Edward M. Riseman. Optimal Geometric Model Matching Under Full 3D Perspective. In Second CAD-Based Vision Workshop, pages 54 -- 63. IEEE Computer Society Press, February 1994. (Submitted to CVGIP-IU).

....sensor, target and scene geometry. This may be thought of as model based sensor fusion, and contrasts with more traditional approaches which attempt to fuse data based upon low level queues only [EG92] The roots of our approach lie in past alignment based object recognition research [Low85, HU90, BR95] which has demonstrated the value of algorithms which precisely vary 3D object to sensor alignment as part recognition. While this paradigm is dominant in many domains, it is surprisingly absent from work on ATR. Instead, ATR is dominated by systems which employ fixed sets of image space ....

J. Ross Beveridge and Edward M. Riseman. Optimal Geometric Model Matching Under Full 3D Perspective. Computer Vision and Image Understanding, 61(3):351 -- 364, 1995. (short version in IEEE Second CAD-Based Vision Workshop).

....by the Advanced Research Projects Agency (ARPA) under grant DAAH04 93 G 422, monitored by the U. S. Army Research Office. Least squares techniques are sensitive to outliers. Median filtering more robustly estimates the true coregistration and is compared with a local search matching capability [3, 4]. Results are presented for both controlled synthetic and real data. 2 Background Motivation A long tradition of work on object recognition has emphasized finding matches between object and image features such that there is a single globally consistent alignment of features. Lowe [20] ....

....controlled synthetic and real data. 2 Background Motivation A long tradition of work on object recognition has emphasized finding matches between object and image features such that there is a single globally consistent alignment of features. Lowe [20] Huttenlocher [15] Grimson [10] and we [5, 6, 3, 4] have all proposed system within this paradigm. Our work here provides an example of extending these methods to multiple constrained sensors. Others have worked on problems similar to the coregistration problem discussed in this paper. Herbert [11] presents a least squares mechanism for computing ....

J. Ross Beveridge and Edward M. Riseman. Optimal Geometric Model Matching Under Full 3D Perspective. Computer Vision and Image Understanding, 61(3):351 -- 364, 1995. (short version in IEEE Second CAD-Based Vision Workshop).

....match the outline of a truck to a subset of line segments extracted from a photograph in which a truck appears. The field of line segments may also include clutter and the matched object may be fragmented or distorted. This is a difficult problem with important practical applications [BR95, Bev93, CB93, FHR 90] It is therefore interesting to see whether genetic algorithms can improve performance beyond the baseline established using local search. Several important themes emerge from our studies. The first is the importance of distinguishing problems with strong nonlinear ....

....ascent that uses a rapid heuristic approximate evaluation method to filter neighbors from further consideration. This reduces the cost of evaluating all L neighbors during one step of steepest ascent. 5 Near Optimal Geometric Matching 5. 1 Background For several years Beveridge [BWR91] Bev93] BR95] has studied the use of local search as a way of finding near optimal solutions to geometric matching problems. Here, geometric matching is the problem of solving for the optimal correspondence mapping and geometric transformation which relate a model to a set of data. For example, the model may ....

[Article contains additional citation context not shown here]

J. Ross Beveridge and Edward M. Riseman. Optimal Geometric Model Matching Under Full 3D Perspective. CVGIP: Image Understanding, page (to appear) , 1995. (short version in IEEE Second CAD-Based Vision Workshop).

....Research Office get and scene geometry. This may be thought of as model based sensor fusion, and contrasts with more traditional approaches that attempt to fuse data based upon low level cues only [ 20 ] The roots of our approach lie in past alignmentbased object recognition research [ 41; 31; 8 ] In this line of research, the value of varying 3 D object to sensor alignment during recognition has been clearly demonstrated. While this paradigm is popular in many domains, it is surprisingly absent from work on ATR. Instead, ATR has been dominated by systems which employ fixed sets of ....

....3 5 micron IR images at the Colorado Demo C site in July, the MDT target detection system was selected to run in conjunction with IR based target detection as part of the RSTA package. After a significant software integration effort, the MDT system was finally integrated and debugged on board SSV B in June, 1995. On Tuesday, June 13, a handfull of training images were collected using this vehicle, and the following morning 14 of these images (3 indicating typical background colors and 11 showing vehicles) were used to train a color look up table (LUT) Using this LUT, the system was tested from 1 to 5 PM ....

[Article contains additional citation context not shown here]

J. Ross Beveridge and Edward M. Riseman. Optimal Geometric Model Matching Under Full 3D Perspective. Computer Vision and Image Understanding, 61(3):351 -- 364, 1995. (short version in IEEE Second CAD-Based Vision Workshop).

....line. Traditional methods for locating objects in optical imagery typically use edge detection algorithms. Local edges [34, 20] may be grouped into larger features such as straight line segments [9, 31] These linear features, in turn, may be matched to linear features of stored object models [29, 22, 18, 5]. These bottom up feature extraction algorithms are prone to error [13, 3] and often produce extraneous line segments, fragmented segments, and sometimes over grouped segments. Our past work in other problem domains [2] demonstrated that local search, coupled with sound and efficient tests of ....

J. Ross Beveridge and Edward M. Riseman. Optimal Geometric Model Matching Under Full 3D Perspective. Computer Vision and Image Understanding, 61(3):351 -- 364, 1995. (short version in IEEE Second CAD-Based Vision Workshop).

....our data set, and then conclusions and future work are discussed. 2 Relation to Prior Work Our work draws upon the sensor fusion and CAD based recognition literature. Most uses of CAD based recognition focus upon a single sensor. For example, CAD models have been used for matching to 2D imagery [14, 35, 9], 3D range data [3, 4] as well as multispectral imagery such as IR [29] and SAR [11] Typically these CAD systems rely on either the 3D or 2D geometry of the model to constrain the location and appearance of that object in the imagery [25, 19, 20] With respect to sensor fusion, Aggarwal [1] ....

J. Ross Beveridge and Edward M. Riseman. Optimal Geometric Model Matching Under Full 3D Perspective. In Second CAD-Based Vision Workshop, pages 54 -- 63. IEEE Computer Society Press, February 1994. (Submitted to CVGIP-IU).

....existed for testing our new LADAR, FLIR and color based approach. The result of this data collection effort is over 400 LADAR, FLIR and color images of four different military vehicles. Vehicles are both out in the open and terrain occluded. This activity is described further in Section 5. 1 and [ Beveridge et al. 1994 ] Another major activity is to explore the use of color as an alternative to FLIR for daytime target detection. Results on Fort Carson data show that multivariate decision tree learning techniques can discriminate camouflage from outwardly similar terrain. The current work uses non parametric ....

....to 2D image data subject to a single best fit 2D similarity transformation. Later versions utilized the 3D sensor pose work of Kumar [ Kumar, 1989; Kumar, 1992; Kumar and Hanson, 1994 ] to perform fitting of 3D object models to corresponding features in a 2D image [ Beveridge and Riseman, 1992; Beveridge and Riseman, 1994 ] Currently the work of Kumar is being extended to handle both registration between multiple sensors as well as 3D pose between the sensors and object model. The resulting least squares fitting procedure is what we have chosen to call coregistration . Coregistration is summarized in Section ....

[Article contains additional citation context not shown here]

J. Ross Beveridge and Edward M. Riseman. Optimal Geometric Model Matching Under Full 3D Perspective. In Second CADBased Vision Workshop, pages 54 -- 63. IEEE Computer Society Press, February 1994. (Submitted to CVGIP-IU).

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