Liu, H., Hong, T.H., Herman, M., Camus, T. and Chellappa, R., Accuracy vs Efficiency Trade-offs in Optical Flow Algorithms, Computer Vision and Image Understanding, Vol. 72, No. 3, pp. 271-286, 1998.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....The first case cannot be detected and is removed by assumption. The second and the third can in fact be detected. Different flow algorithms provide different approaches for resolving this problem. Two comprehensive papers on the subject of optical flow performance exist. Work by Liu et al. [39] has studied the efficiency accuracy tradeoff of different algorithms. The authors produce curves of accuracy versus efficiency for comparing different optical flow algorithms. A curve is constructed for each algorithm by changing 13 its parameters. Barron et al. have produced a paper that ....

...., 1 = 15 This approach can be described as a weighted minimization of normal constraints where the weights are the magnitude of the spatial gradient of the image. Barron et al. report in their survey [8] that Lucas and Kanade s algorithm provides the second most accurate results. Liu et al. [39] evaluate Lucas and Kanade as providing the third best efficiency accuracy curve. Thus, there is noted interest in this approach. However, there is an important disadvantage to this algorithm. As discrete temporal differentiation is necessary, strong temporal support is required. The Barron et al. ....

[Article contains additional citation context not shown here]

Liu, H., Hong, T.H., Herman, M., Camus, T. and Chellappa, R., Accuracy vs Efficiency Trade-offs in Optical Flow Algorithms, Computer Vision and Image Understanding, Vol. 72, No. 3, pp. 271-286, 1998.

....to find the vector field v, given the image sequence f (x, y, t ) Motion estimation is a widely studied and applied problem. Numerous researchers have developed diverse methods and several survey papers discuss the relative merits of the various leading methods and compare their performances [3, 7, 16, 23]. In this paper, we are concerned with estimating vector fields v that are parameterized by an affine model. Namely, the vector fields of interest are characterized by y # , 2) where v 0x v 0y # , 3) is a constant vector representing global translational motion, and ab cd ....

....2. Lab. Picture from a webcam at the researchers of fice. The webcam was rotated about 45# so as to create an image in which the majority of image texture is not aligned at 0# and 90# . The image is 240320 pixels. In addition to our own synthetic image sequences, we follow the papers of [3, 16] and measure performance on a well known set of benchmark image sequences from [3] While these image sequences contain many frames, we limit the image sequences to only five frames. In practice, this represents a reasonable number of frames as often in real image sequences the vector field v ....

[Article contains additional citation context not shown here]

Hongche Liu, Tsai-Hong Hong, M. Herman, T. Camus, and R. Chellappa, "Accuracy vs. efficiency trade-offs in optical flow algorithms," Computer Vision and Image Understanding, Vol. 72, pp. 271--286, 1998.

....valid and useful for many applications. Acknowledgments The authors would like to acknowledge the broadcasting station TV Noord for their help in recording the video data. 3 An initial discussion about the trade off of the accuracy of optic flow algorithms and their efficiency can be found in [21]. 14 ....

H. C. Liu et al., "Accuracy vs. efficiency trade-offs in optical flow algorithms," Computer Vision and Image Understanding, vol. 72, pp. 271--286, Dec. 1998. 15

....hardware, then the run time performance of even complicated algorithms may be analyzed using standard software tools and visualization techniques, without having to fight the intricacies of some specialpurpose image processor. Many techniques for optical flow exist (e.g. 2, 8, 9] see also [10, 11, 12] for reviews and discussions of several techniques) Although many of these techniques can perform very well for certain sequences of images, there are very few that are currently able to support real time performance without special purpose hardware. One obvious reason calculating optical flow is ....

....would be too computationally expensive, a gradient based search in multiresolution image and parameter spaces is performed, with the initial coarser resolutions helping to avoid local minima. Using an 80 MHz HyperSPARC computer an execution time of 123 seconds is estimated for 100 dense results [12]. A faster algorithm yielding much coarser results (similar to the sparse spatial sampling of [31] would result in an algorithm of about a couple of seconds on an 80 MHz HyperSPARC. Accuracy results are not reported for this faster algorithm, but would depend on the effectivness of the affine ....

[Article contains additional citation context not shown here]

H. Liu, T. Hong, M. Herman, R. Chellapa, "Accuracy vs. Efficiency Trade-offs in Optical Flow Algorithms", Proceedings of the Fourth European Conference on Computer Vision, Cambridge England, April 1996

....ffl Cost : The cost at which estimates are produced. This is crucial in computer vision e.g. in active vision, where there are time constraints to be met. We discuss the notion of cost , as it may have many interpretations. Only recently have these two concepts been studied jointly 1 , e. g in [9]. To our knowledge, there are very few results linking the quality of an estimator to the computational 1 Experimental design , or experience planning , goes in that direction, by considering the cost at which one does observations. However, it usually considers the cost of doing ....

H. Liu, T. Hong, M. Herman and R. Chellappa. "Accuracy vs. efficiency trade-offs in optical flow algorithms". In Proc. of the 4th European Conference on computer Vision, Vol. 2, pp. 174-183, Cambridge, UK, 1996.

....are given the same computational resource, and we discuss how being able to build estimators of a given computational cost can help in maximizing the efficiency. An interesting study of how the performance of optic flow estimators varies with the available computational resource is given in in [8]. As opposed to that paper, we do not consider procedures that compute the optic flow in every pixel, but the affine flow (the flow and its first derivatives) After a brief introduction to optic flow, in Section 2, we present our approach in Section 3, detailing the points where it differs from ....

....(computationally) unless new theoretic tools help us improve its efficiency. Being able to compare algorithms and parameter setups on specific data does not give us much insight on the reasons of their performance. An interesting study of performance versus time cost appeared recently in [8]. In our case, this is equivalent to performance versus number of estimates (when all other parameters are fixed) In turn, one can show theoretically [6] and observe in practice that the error is inversely proportional to the number of observations. Still more interesting would be a study ....

Hongche Liu, Tsai-Hong Hong, Martin Herman and Rama Chellappa, "Accuracy vs. efficiency trade-offs in optical flow algorithms", Proc. of the 4th ECCV, vol. 2, pp. 174-183, Cambridge, UK, 1996.

....phase of the CDWT coefficients is invariant to both kinds of perturbation. This gives them a significant advantage over the intensity based methods, which are clearly very sensitive even to small uniform perturbations to intensity. 15 D. Noise immunity For this test, which was suggested by Liu [23], white Gaussian noise of varying standard deviation was added to both frames in each of the synthetic sequences. Figure 13 contains the plots of mean error angle against noise standard deviation for the two CDWT algorithms and the gradient based pel recursive algorithm. The FBM results are ....

....clearly in the case of the Yosemite sequence (Figure 13(c) because the discontinuity at the horizon dominates the error. Note that in general the rise in error is approximately linear with noise. This contrasts with the quadratic noise dependence of the phase based algorithm of Fleet and Jepson [23]. E. Video coding Measuring performance on real sequences is more difficult as one rarely has access to the true motion field, even if it exists in the sense described above (which ignores the common phenomena of occlusion and uncovering. Instead one must make use of simple simulated video ....

H. Liu, T-H. Hong, M. Herman, and R. Chellappa, "Accuracy vs efficiency trade-offs in optical flow algorithms," in Proc. Fourth European Conference on Computer Vision. April 1996, vol. II, pp. 174--183, Springer-Verlag.

....problematic in our case where a wide angle lens is used and individual objects occupy only a small fraction of the visual field. Calculating a least squares best fit to the correlation surface as in [3] was ruled out due to real time performance requirements (see efficiency experiments in [22][24]) In order to satisfy our real time requirements, a fast approximation method was used to derive continuously valued flow fields. To avoid a computationally expensive search for the true flow, the 2 dimensional interpolation is decomposed into two 1 dimensional interpolations; the first ....

H. Liu, T.-H. Hong, M. Herman, R. Chellapa, "Accuracy vs. Efficiency Trade-offs in Optical Flow Algorithms", Pro- ceedings of the Fourth European Conference on Computer Vision, Cambridge, England, April 1996.

....optical flow can be computed on 32x64 images, quantizing flow speeds into 5 bins, at 35 frames per second on an 80 MHz Themis HyperSPARC 10 computer. Although this algorithm is not as accurate as many other optical flow algorithms, it is generally superior in terms of computational efficiency [23]. In our implementation, an original image of 256x512 pixels, captured with a field of view camera, is subsampled to 32x64 pixels using a simple block subsampling algorithm which averages an NxN block of pixels. This simple subsampling is very fast and (a) The search area for the example pixel ....

H. Liu, T.-H. Hong, M. Herman, T. Camus, R. Chellapa, "Accuracy vs. Efficiency Trade-offs in Optical Flow Algo- 39 rithms", Computer Vision and Image Understanding, Vol. 72, No. 3, pp. 271-286, Academic Press, December 1998.

....[4] and Camus [6] is Figure 6: Allocation of independently moving object functions to hardware resources. Compute Independent Motion Filters SPARC Video A D Compute Flow MV 200 Overlay Video D A Filter Independent Motion independent motion Monitor camera 11 depicted in Figure 10 (see [17]) This figure shows accuracy (or error) as one coordinate and efficiency (or execution time) as the other. Two dimensional accuracy efficiency (AE) curves in this figure characterize an algorithm s performance. A curve is generated by setting parameters in the algorithm to different values. For ....

Liu, H., Hong, T., Herman, M. and Chellappa, R., "Accuracy vs. Efficiency Trade-offs in Optical Flow Algorithms ", Proceedings of the Fourth European Conference on Computer Vision, Cambridge, England, 1996.

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