R. Dutta, R. Manmatha, L.R. Williams, and E.M. Riseman. A data set for quantitative motion analysis. CVPR, 1989.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....update on the predicted state vector s(k=k Gamma 1) The system is then ready to process the next image in the sequence. 5 Experimental Results The estimation algorithms developed in the earlier sections were tested on a real image sequence.The sequence used is the UMASS rocket al..V sequence [4]. The first, eighth and final images of the real image sequence (consisting of 16 images 1 ) are shown in Fig. 1(a) c) The results of feature extraction are shown in Fig. 1(d) Image plane trajectories of the 11 selected features are shown in Fig. 1(e) and (f) Four points out of the 11 are ....

R. Dutta et al., "A data set for quantitative motion analysis," in IEEE Computer Society Conf. on Computer Vision and Pattern Recognition, (San Diego, CA), pp. 159--164, June 1989.

....of the estimated epipole e, with jej = 60. Figures 1 and 2 show the image plane and directional errors on the grid e x = 5 : 1 : 5; e y = 5 : 1 : 5 (in MATLAB notation) for sequences created using structures from the PUMA and UMASS Martin Marietta Rocket Field real image sequences [14][3]. The figures show the results of minimizing over the rotation and structure for fixed e. The angle between the length 121 vectors of directional and image plane errors was 0:35 . for the PUMA and 1:2 and for the Rocket sequence. The PUMA sequence had e = 2:5; 9:7) jTj=Z min = 0:3, ....

....minimization over the structure unknowns but still compute the exact error. We report results for noiseless synthetic sequences generated using the measured ground truth structure from two real motion sequences: the UMASS PUMA sequence [14] and the UMASS Martin Marietta Rocket Field sequence [3]. The depths for the PUMA structure vary from 13 32 and for the Rocket Field structure they vary from 17 to 67. We used several translations ranging in magnitude up to 7:2 for the PUMA structure and up to 4:2 for the Rocket Field structure. For simplicity, we took the true rotation to be zero; ....

[Article contains additional citation context not shown here]

R. Dutta, R. Manmatha, L.R. Williams, and E.M. Riseman, "A data set for quantitative motion analysis," CVPR, 159-164, 1989.

....to within Y of its true value. Frame 1 Frame 9 Frame 16 Figure 2: Example frames from the real image sequence. Tracked features are shown in white. 4. 1 Experiments on Real Imagery: Box sequence In this section, we describe experiments with the BOX sequence available from the UMass database [5]. Fig. 2 shows example frames from the sequence. Corner features were tracked using an implementation of the Kanade Lucas Tomasi feature tracker available from http: vision.stanford.edu # birch klt . The corner features on the front face of the box were tracked and used as measurement input to ....

R. Dutta, R. Manmatha, L.R. Williams, and E.M. Riseman. A data set for quantitative motion analysis. CVPR, 1989.

....80 lines of MATLAB code, and Steps P4 and P5 require an additional 30. The iterative version of the Sturm Triggs approach requires about 80 lines. We created synthetic sequences using the ground truth structure from two real image sequences: the UMASS Martin Marietta rocket field sequence [3] and the UMASS PUMA sequence [12] The points in the rocket field sequence range from 17 to 67 in depth and cover an effective FOV of 37 ffi , while the PUMA points range from 13 to 32 in depth and cover an effective FOV of 33 ffi . Table 2 gives the parameters of the experiments, and Table 3 ....

....10 Gamma9 . On average, Step L5 converged in 7.5 cycles and always in less than 15. Figure 5 shows similar results for the homography recovery, for the homography error defined in Section 3.4. We also tested Algorithm II and the Sturm Triggs approach on the rocket field real image sequence [3], see Figure 4. This sequence has large translations ranging up to 7:5 in magnitude (recall that the depths vary from 17 to 67) The camera moves approximately along a line, with T i deviating from its average direction by up to 1:5 ffi . Steps L2 and L3 of our approach recovered T 0 with ....

R. Dutta, R. Manmatha, L.R. Williams, and E.M. Riseman, "A data set for quantitative motion analysis," CVPR, 159-164, 1989.

....we avoid a costly minimization over the structure unknowns but still obtain exact, tight bounds on the true error. We report results for noiseless synthetic sequences generated using the measured ground truth structure from two real motion sequences: the UMASS Martin Marietta rocket eld sequence [5] and the UMASS puma sequence [17] The depths for the Rocket Field sequence vary from 17 to 67, and for the PUMA sequence they vary from 13 32. We use a variety of translations and, for simplicity, take the rotation to be zero. We set the true rotation to be zero; the error surfaces in the gures ....

R. Dutta, R. Manmatha, L.R. Williams, and E.M. Riseman, \A data set for quantitative motion analysis," CVPR, 159-164, 1989.

....within 0:5 o of its true value. Frame 1 Frame 9 Frame 16 Figure 2: Example frames from the real image sequence. Tracked features are shown in white. 4. 1 Experiments on Real Imagery: Box sequence In this section, we describe experiments with the BOX sequence available from the UMass database [5]. Fig. 2 shows example frames from the sequence. Corner features were tracked using an implementation of the Kanade Lucas Tomasi feature tracker available from http: vision.stanford.edu # birch klt . The corner features on the front face of the box were tracked and used as measurement input to ....

R. Dutta, R. Manmatha, L.R. Williams, and E.M. Riseman. A data set for quantitative motion analysis. CVPR, 1989.

....that we place no contraint on the rotations and do not require the magnitudes of the translations to be constant; only the translation direction is assumed approximately constant. In our experimental section, we report results for a real image sequence the Martin Marietta Rocket Field sequence [1, 16] for which actually both the translation direction and translation magnitudes vary. We have argued previously [9, 10] that there probably is no one SFM algorithm producing 1 Our approach can be implemented in recursive or batch mode; this is a separate issue from the use of noise sensitive ....

....yields no improvement in the reconstructions. Though the use of this approximate procedure is heuristic compared to the rigorously justifiable multiplicative iteration, its use may actually give better results. 3 Experiments: Rocket Sequence Our experiments were carried out on the Rocket sequence [1, 16], using 22 feature point correspondences provided to us by J. Thomas. Ground truth was available for only 11 of these; depth reconstructions are reported only for these points. For this sequence the camera rotations are small, averaging 2.2 degrees with respect to the first camera position and ....

R. Dutta, R. Manmatha, L.R. Williams, and E.M. Riseman, "A data set for quantitative motion analysis," CVPR, 159-164, 1989.

....broadly [26, 24, 4] often cases intermediate between two domains can be handled by approaches specialized for either. For instance, the algorithms of [30] and [27] respectively, domain 3 style and domain 1 style approaches) have both worked on the Martin Marietta rocket field sequence [6] 33 . Second, if a sequence spans the extremes of different domains, admittedly none of the algorithms suggested above may work on the entire sequence. But it is always possible to select parts of the sequence that do belong to single domains 34 . One can apply specialized algorithms to ....

R. Dutta, R. Manmatha, L.R. Williams, and E.M. Riseman, "A data set for quantitative motion analysis," CVPR, 159-164, 1989.

....experiment, the rocket field image sequence from the IEEE motion workshop database was used. This is an outdoor sequence produced using a camera mounted on an autonomous vehicle; the first image is displayed in Figure 2. Ground truth was available for 12 of the 15 points tracked over 11 frames [4]. Points were mostly tracked by hand but some were obtained by running Anandan s optical flow algorithm [1] The robot motion was a constant forward translation (in steps of about 0.9 m) and very small rotations. In Table 4.1 the average 3D distance error for 10 of the 12 tracked points is ....

R.Dutta, R.Manmatha and L.R.Williams, "A Data Set for Quantitative Motion Analysis," CVPR, San Diego, California, pp. 159-164, 1989.

....direction. For these cases the algorithm presented here must be modified. We have already presented results for real and synthetic sequences using appropriately modified algorithms in [16, 15] The real image sequences used were the UMASS PUMA [20] and the Martin Marietta Rocket Field sequence [3]. We have also shown experimentally that it is possible to distinguish which method is appropriate directly from the data. For instance, it is easy to determine that the translational motion is not general if DCH has fewer than three leading singular values. In this paper, we report the testing of ....

R. Dutta, R. Manmatha, L.R. Williams, and E.M. Riseman, "A data set for quantitative motion analysis," CVPR, 159-164, 1989.

....of the center of mass, b) bouncing velocity, c) instantaneous angular velocity, and (d) quaternions. Eight feature points were used in both models. Instead of randomly generating these feature points, we used feature points from the Rocket al..V sequence, in which ground truth is available [6, 15]. Note that in this above sequence, the camera moves along a straight path toward the left. In our experiment, we shifted each point upward and to the right by 10.0 so that the corresponding image point would be close to the center of the image. To avoid the depth becoming negative after a few ....

R. Dutta et al. A data set for quantitative motion analysis. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition, pages 159--164, San Diego, CA, June 1989.

....that compute structure and motion with no initial guess. A class of algorithms (batch and recursive) is developed that accomplishes this, where the appropriate algorithm depends on the particular image sequence. Experiments are described on the PUMA sequence [13] and the Rocket Field sequence [2]. 1 Introduction The approach to multiframe structure from motion (MFSFM) for point features described here may be seen as a generalization of the earlier work by Tomasi [16] to the case of full perspective for instance, our approach works well on the Rocket Field sequence, where perspective ....

R. Dutta, R. Manmatha, L.R. Williams, and E.M. Riseman, "A data set for quantitative motion analysis," CVPR, 159-164, 1989.

....corrupted by spatial or temporal aliasing. This enables us to test basic implementations of differential methods and matching methods on the same data without the complexities of hierarchical coarsefine control and warping techniques. For example, we do not consider stop and shoot sequences [18]. Barron, Fleet and Beauchemin IJCV 12:1, pp43 77, 1994 3 This paper concentrates on the accuracy and density of velocity estimates produced by the nine methods. Confidence measures have been used to extract subsets of estimates for which we report error statistics. While confidence measures are ....

Dutta, R., Manmatha, R., Williams, L., and Riseman, E.M. (1989) A data set for quantitative motion analysis. Proc. IEEE CVPR, San Diego, pp. 159-164

....ffi . An additional set of 100 trials where the plane normal was allowed to vary randomly near the set of image plane directions produced similar results. These results are consistent with those reported in Table 2. We also report results on two real image sequences: the rocket field sequence [1, 12] and the PUMA sequence [9, 10] The rocket sequence consists of nine images obtained by approximately forward motion in an outdoor environment. The maximum translation magnitude, between the first and ninth image, was 7.35 feet, while the depth of the nearest point relative to the first (farthest) ....

R. Dutta, R. Manmatha, L.R. Williams, and E.M. Riseman, "A data set for quantitative motion analysis," CVPR, 159-164, 1989.

....error depends on the closeness of image motion to an integer number of pixels. Hierarchical differential based methods (using image warping or registration) may provide an alternative to correlation methods. 4 Dutta et al. s stop and shoot sequences constitute interesting image sequence examples [42]. ACM Computing Surveys, Vol. 27, No. 3, pp. 433 467, 1995 11 One of the purposes of this study [14] was to analyze the performance of different optical flow methods and to and to encourage others to compare numerical results with theirs. Towards this end, several authors now compare the ....

R. Dutta, R. Manmatha, L. Williams, and E. M. Riseman. A data set for quantitative motion analysis. In IEEE Proceedings of CVPR, pages 159--164, San Diego, California, June 1989.

....The sequence, with correspondences and true 3D data, is available from the computer vision group at the University of Massachusetts at Amherst. Hence, one can measure the performance of reconstruction at a quantitative level. Reconstruction results on this set by other approaches can be found in [6, 34, 23, 46, 1]. On the sneaker images, a set of points were manually selected on one of the frames, referred to as the first frame, and their correspondences were automatically obtained along all other frames used in this experiment (corresponding points are marked by overlapping squares in Fig. 5) The ....

....constraints such as coplanarity, perpendicular lines and planes in the scene, and known distances between points in the scene (the latter constraint is non linear) Since the true 3D data is provided with the sequence we chose the former method. To compare our results with those reported in [6, 34, 23, 46, 1], the level of error was calculated as the ratio of the average mean square error in depth with the overall average depth of the cube. The average depth values of the sample points on the cube is 626.48 units. The average error we found between the the given depth values and the reconstructed ....

[Article contains additional citation context not shown here]

R. Dutta, R. Manmatha, L. R. Williams, and E. M. Riseman. A data set for quantitative motion analysis. In 1989 IEEE Conference on Computer Vision and Pattern Recognition, Los Alamitos, CA, June 1989. IEEE Computer Society, IEEE Computer Society Press. (San Diego, CA).

....outdoor scenes, we are enhancing the system to support the generation of fractal terrain models, plant shapes, and corresponding textures. 3. 5 Ground Truth Data The acquisition of ground truth data is usually expensive and tedious (e.g. through separate theodolite or radar measurements [7, 13]) therefore the pos sibility to access highly reliable ground truth data is a key motive for using simulation in many vision applications. Rendering systems like Radiance do not have this access capability built in, but most of the internal data structures are already available and Radiance is ....

R. Dutta, R. Manmatha, L R. Williams, and E.M. Riseman. A data set for quantitative motion analysis. In Proc. Conf. on Computer Vision and Pattern Recognition, pages 159--164, June 1989.

....estimate 0 20 40 60 80 100 0 10 20 30 40 50 Two frame estimate Figure 8: Left: translation direction errors for 800 trials with approximately planar scenes. Step translation size 0:2 with 10 percent added random translations, 20 Slant 70, F = 60. Right: One frame of rocket sequence [2]. reconstruction on these planar sequences with approximately linear motion. We repeated the experiment with exactly linear motion and found average errors of 4:7 and 5:1 for min i Theta i T; T G j ; Theta i T; T A jj (our algorithm) and min i Theta i T 2 ; T G j ; Theta ....

....i T; T G j ; Theta i T; T A jj while the two frame algorithm again gave 5:1. Thus the major problem our algorithm faces for planar scenes is the poor initial rotation compensation, and this can be corrected by recompensating. Real Image Sequences. The Martin Marietta Rocket sequence [2] (Figure 8 (right) consists of 9 images of 22 tracked points. The scene is approximately a plane slanted at 84 ffi , where the 3D points deviate from the plane on average by 0:9. Also, 17 Z 61, and the FOV occupied by the image points was F = 25 ffi . The nominal FOV was (72 ffi ; 57 ....

[Article contains additional citation context not shown here]

R. Dutta, R. Manmatha, L.R. Williams, and E.M. Riseman, "A data set for quantitative motion analysis," CVPR, 159-164, 1989.

.... Note that since the image displacement, I u I, is necessarily much less than the focal length of the camera (recall that f is typically several hundred pixels) even a small error in 12 SVehicle (not camera) rotations measured by a land nav igation system on the Autonomous Land Vehicle (ALV) [17] strongly suggest a bell shaped distribution for the rotation R of the ALV around the axis perpendicular to its base when it is moved approximately straight ahead. Since the ALV is 16000 pounds in weight, eight wheel powered, and hydrostatically driven, it is expected that the spread of the angles ....

R. Dutta, R. Manmatha, L. R. Williams, and E. M. Rise- man. A data set for quantitative motion analysis. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 159164, June 1989.

....it can be computed from the table that with an absolute error in rotation of 0.100 for displacement vectors of 5 pixel length in 45 out of 100 cases the relative error in depth will be more than 10 6 4R, otations measured by a land navigation system on the Au t . Land Vehicle (ALV) [9] strongly suggest a bell shaped distribution for the rotation R of the ALV around the axis per pendicular to its base when it is moved approximately straight ahead. Actual data from a 30 frame sequence collected on almost level ground has the spread of R at about 2.4 . 5R, esults similar in ....

P. Dutta, P. Manmatha, L. P. Williams, and E. M. Piseman. A data set for quantitative motion analysis. Proceedings of the IEEE Computer Society Conference on Computer Vision and Patter tecognition pages 159 164 June 1989.

....cone 2, cone 3 and cone 4 (as shown in Figure 2) are standing out clearly from the ground plane in the reconstructed 3 D surface. Autonomous Land Vehicle sequence A second experiment was done on a sequence collected via the Autonomous Land Vehicle. The data collection process is detailed in [9]. Results on this sequence are shown in Figure 6. The distant mountain is clearly identified. Umass Denning Robot sequence A third experiment was done on a sequence collected via the Denning robot at the University of Massachusetts at Amherst. This image was taken indoor under poor lighting ....

R. Dutta, R. Manmatha, L. R. Williams, and E. M. Riseman. A data set for quantitative motion analysis. IEEE Computer Society Conference on Computer Vision and Pattern Recognition, June 1989.

....generalizes from the work of Heeger and Jepson [1992] on recovering translational motion from optical flow; it gives an effective method for determining the translation from sparse as well as dense optical flow. The approach has been tested on two real image sequences: the Rocket Field sequence [Dutta 1989], which involves straightahead translation and significant perspective effects, and the difficult PUMA sequence [Sawhney 1990] Fast and accurate reconstructions were obtained for both these sequences on a DecStation 5000 using MATLAB. The approach is based on finding suitable approximations of ....

Dutta, R., Manmatha, R., Williams, L., and Riseman, E., "A data set for quantitative motion analysis," Proc. CVPR, San Diego, CA, June 4-8, 1989, 159-164.

No context found.

R. Dutta, R. Manmatha, L.R. Williams, and E.M. Riseman. A data set for quantitative motion analysis. CVPR, 1989.

No context found.

R. Dutta, R. Manmatha, L.R. Williams, and E.M. Riseman. A data set for quantitative motion analysis. CVPR, 1989.

No context found.

R. Dutta, R. Manmatha, L R. Williams, and E.M. Riseman. A data set for quantitative motion analysis. In Proc. Conf. on Computer Vision and Pattern Recognition, pages 159#164, June 1989.

First 50 documents

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