R. Szeliski, Sing Bing Kang, and HeungYeung Shum. A parallel feature tracker for extended image sequences. In Proc. International pages 241--246, 1995.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....correspondence is a central problem in computer vision. A variety of interactive tools rely on a human operator to match features across images [BB95, DTM96, SHS98] These tools do not scale e#ectively, and are vulnerable to operator error and numerical instability. Low level feature trackers [SKS95] and texture trackers [HS81, ZMI00] recover correspondence under short baseline motion (i.e. constant brightness and little or no occlusion) Several robust statistical techniques randomly generate match sets or aligning transformations, selecting the most consistent [FB81, CC91, FZ98, Ste98, ....

R. Szeliski, Sing Bing Kang, and HeungYeung Shum. A parallel feature tracker for extended image sequences. In Proc. International pages 241--246, 1995.

....only up to an arbitrary projective transformation. Other linearized versions of SFM have been formulated, based on SVD [41] or a#ne approximations [33] 5.4 Correspondence Methods Nearly all registration algorithms rely on explicit knowledge of correspondence between features. Low level trackers [50] and dense texture trackers [31, 58] attempt to compute correspondence under short or infinitesimal baselines (i.e. for situations in which scene brightness and viewpoint change little across images, and there is little or no occlusion) Robust statistical techniques have been developed to ....

R. Szeliski, Sing Bing Kang, and Heung-Yeung Shum. A parallel feature tracker for extended image sequences. In Proc. International Symposium on Computer Vision, pages 241--246, 1995.

....strategy is important as it accounts for not only occlusions, but also perspective distortion due to changes in object pose. In contrast to Lowe s approach [13] which uses edges, we use whole predicted images. 2. 1 Tracking by spline based registration In the spline based registration framework [22, 24], a new image I 2 is registered to an initial base image I 1 using a sum of squared differences formula E(fu i ; v i g) X i [I 2 (x i u i ; y i v i ) Gamma I 1 (x i ; y i ) 2 ; 1) where the fu i ; v i g s are the per pixel flow estimates. In this registration technique, the flow ....

....min;ij (12) This is particularly important in the case of face model recovery because of the possible lack of texture on parts of the face, such as the cheeks and forehead areas. Using this metric for c ij downplays the importance of points on these relatively untextured areas (see, for example, [18, 24]) To account for occlusions, c ij is set to zero if the corresponding point is predicted to be hidden. The other term in (8) is E geom (a) X i i ff i jh i Gamma h 0 i j 2 fi i jp i Gamma p 0 i j 2 j ; 13) 2.3 Least squares minimization with geometric constraints 9 which is ....

R. Szeliski, S. B. Kang, and H.-Y. Shum. A parallel feature tracker for extended image sequences. In IEEE International Symposium on Computer Vision, pages 241--246, Coral Gables, Florida, November 1995.

....completely general local motion, as warranted by the data in a given image sequence. By examining the local certainty in the flow computation, we can also use our 2 2 Previous work algorithm as a parallel feature tracker for very long motion sequences where image deformations may be significant [Szeliski et al. 1995]. The motion estimation algorithms developed in this paper can be used in a number of applications. Examples include motion compensation for video compression, the extraction of 3D scene geometry and camera motion, robot navigation, and the registration of multiple images, e.g. for medical ....

....we use patches of varying size, we completely tile the image with patches, and we have no motion discontinuities across patch boundaries. Our motion estimator can be used as a parallel, adaptive feature tracker by selecting spline control vertices with low uncertainty in both motion components [Szeliski et al. 1995]. 3 General problem formulation The general motion estimation problem can be formulated as follows. We are given a sequence of images I t (x; y) which we assume were formed by locally displacing a reference image I(x; y) with horizontal and vertical displacement fields 1 u t (x; y) and v t (x; ....

[Article contains additional citation context not shown here]

R. Szeliski, S. B. Kang, and H.-Y. Shum. A Parallel Feature Tracker for Extended Image Sequences. Technical Report 95/2, Digital Equipment Corporation, Cambridge Research Lab, April 1995.

....This yields estimates of optic flow, which in turn is used by a local tracker to refine the amount of feature displacement. The optic flow between a pair of cylindrical panoramic images is first estimated using splinebased image registration between the pair [Szeliski and Coughlan, 1994; Szeliski et al. 1995] In this image registration approach, the displacement fields u(x; y) and v(x; y) i.e. displacements in the x and y directions as functions of the pixel location) are represented as two dimensional splines controlled by a smaller number of displacement estimates which lie on a coarser ....

....stereo 9 1. the spline based image registration technique is capable of tracking features with larger displacements. This is done through coarse to fine image registration; in our work, we use 6 levels of resolution. While this technique generally results in good tracks (sub pixel accuracy) [Szeliski et al. 1995], poor tracks may result in areas in the vicinity of object occlusions disocclusions. 2. the Shi Tomasi tracker is a local tracker that fails at large displacements. It performs better for a small number of frames and for relatively small displacements, but deteriorates at large numbers of frames ....

[Article contains additional citation context not shown here]

R. Szeliski, S. B. Kang, and H.-Y. Shum. A parallel feature tracker for extended image sequences. In IEEE International Symposium on Computer Vision, Coral Gables, Florida, November 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