D. Reynard, A. Wildenberg, A. Blake, and J. Marchant. Learning dynamics of complex motion from image sequences. ECCV-96, pp. 357--368, Cambridge, UK, 1996. 920  Home/Search   Document Details and Download   Summary   Related Articles   Check

....our choice of models by taking into account one behavior this model must have. As the computation of a cue converges, the entropy (or average uncertainty) of the model should be decreasing, so that a completed computation is una#ected by the model. This rules out stochastic di#erential equations [28, 29] that are often used as motion predictors, as they have constant covariance. The models developed in this section enable existing iterative vision algorithms to be used in the implemented system. The next description describes this implemented system, as well as its specific model for partial ....

Reynard, D., Wildenberg, A., Blake, A., Marchant, J.: Learning dynamics of complex motions from image sequences. In: ECCV '96. (1996) I:357--368

....state of the art as presented in the tutorial. 5. 2 Visual Tracking in Monocular Images Visual tracking has emerged as an important component of systems using vision as feedback for continuous control [63, 64, 67, 66] human computer interfaces [68, 69, 70] surveillance, or visual reconstruction [73, 74, 77, 78]. The central challenge in visual tracking is to determine the image position of an object, or target region of an object, as the object moves through a camera s field of view. This is done by solving what is known as the temporal correspondence problem: the problem of matching the target region ....

D. Reynard, A. Wildenberg, A. Blake, and J. Marchant, "Learning dynamics of complex motions from image sequences," in Proc. European Conf. on Computer Vision, pp. I:357-- 368, 1996.

....) X(t k ) 2) and = 3) The unknown parameters that need to be solved for (and are unique to each signature) are A and B each of which are 2d Theta 2d matrices. There are many methods for learning them, including that of Yule Walker [8] or that used here, due to Reynard et al. [9]. Once A and B are known, a new point in the sequence can be iteratively generated using Equation 1. 3 The Basic Algorithm Here we present the basic algorithm for generating new video clips, based on an example sequence. The first step is to perform principal component analysis on the ....

D. Reynard, A. Wildenberg, A. Blake, and J. Marchant. Learning Dynamics of Complex Motions from Image Sequences. Proc. European Conference on Computer Vision, 1:357--368, 1996.

....of all three peaks. Even though the edge detector described above is quite selective, as the edge segment moves through clutter, we can expect multiple local maxima to appear in the convolution output. This is a well known and unavoidable problem for which many solutions have been proposed [38]. By default, X Vision declares a match if and only if a unique local maximum exists within an interval about the response value stored in the state. The match interval is chosen as a fraction of the difference between the matched response value and its next closest response in the previous frame. ....

D. Reynard, A. Wildenberg, A. Blake, and J. Marchant. Learning dynamics of complex motions from image sequences. pp. 1:357 368, 1996.

.... [3] the authors have proposed to model jointly the random variations in shape arising within from the differences between objects in a given class with those occurring during object motion (due to projective effects) The distinction between the two sources of variability has been pointed out in [6, 7, 22]. In these works, the learning method is rather dynamic, using and modeling temporal image sequences. Example motions are exclusively learned by general purpose tracker based on the assumption that the identity of the object does not change over time. The learned dynamics are then used in a Kalman ....

.... deformation modes associated to the object of interest from real image data without any human interaction (such as manual point 4 correspondence) The method requires that the object be described as the deformations of a single prototype object [12, 2, 18, 3] In contrast to works reported in [6, 7, 22], our approach aims at identifying the shape variability of an object class from temporal image sequences. We assume the whole estimated deformations of one single object enable to constrain the class variability of any objects belonging to the same object class. Our method differs from works ....

D. Reynard, A. Wildenberg, A. Blake and J. Marchant, "Learning dynamics of complex motions from image sequences",. Proc. European Conf. on Computer Vision, Cambridge, UK, April 1996, pp. 357--368.

....motion parameters encode the location of the head, as well as facial displays and expressions. This division is often built into face models [3, 6, 16, 19, 28] to simplify model construction or estimation, and has been used to facilitate learning the variability of motions for a class of objects [25]. The ultimate goal of this separation is to produce an estimation problem with lower dimension. During estimation, the change in the shape parameters should tend to zero as the shape of the observed object is established. Once this occurs, fitting need only continue for the motion parameters. ....

D. Reynard, A. Wildenberg, A. Blake, and J. Marchant. Learning dynamics of complex motions from image sequences. In Proceedings ECCV '96, pages I:357--368, 1996.

....state [3] In these cases, the value of the learned model is eeting, since few targets ever maintain xed dynamics. More interesting learning focuses on models of motion. Existing research includes learning of multi state motion models of targets which exhibit a few discrete patterns of motion [8, 13]. Our work focuses on bootstrap initialization of nonparametric models for the static texture of faces. In contrast with previous work, we explicitly consider issues of automatic learning during tracking without manual intervention. 4 Nonparametric Texture Models In our rst example, we ....

D. Reynard, A. Wildenberg, A. Blake, and J. Marchant. Learning dynamics of complex motions from image sequences. In ##### ######## ##### ## ######## ######, pages 357-368, 1996.

....a tracking session. This separation produces an easier tracking problem by requiring a smaller description of object state to be estimated in each frame. This division is often built into face models [BY95, LRF93, MRB95, TW93] to simplify model construction or estimation, while Reynard, et al. RWBM96] use this separation to permit learning the variability of motions for a class of objects. As a result of this separation, the parameters in q are rearranged and separated into q b , which describe the underlying features of an individual, and into qm , which describe motion (both rigid and ....

D. Reynard, A. Wildenberg, A. Blake, and J. Marchant. Learning dynamics of complex motions from image sequences. In Proceedings ECCV '96, pages I:357--368, 1996.

....roller blading, and cycling. Finally, the system s noise stability properties have been evaluated using synthetic data sets, and results are encouraging. 2 Related Work The extended Kalman filter (EKF) has proven to be very useful in recovery of rigid motion and structure from image sequences [7, 1, 5, 22, 20, 24]. Most of these approaches assume rigid motion. One of the first important results on recursive structure and motion estimation was the work of [7] The formulation of [1] yields improved stability and accuracy of the estimates. In both methods, image feature tracking and correspondence are ....

.... , is the process noise and ( the system evolution matrix, is based on first order Newtonian dynamics in 3D space and assumed time invariant ( 0 . If additional prior information on dynamics is available, then can be changed to better describe the system evolution [22]. Our measurement vector is 1 ( 2 ( where 43 3 are the image plane coordinates for the observed feature 5 at time . The measurement vector is related to the state vector via the measurement equation: 1 ( 76 ( 98 ( Note ....

D. Reynard, A. Wildenberg, A. Blake, and J. Marchant. Learning dynamics of complex motions from image sequences. ECCV, 1996.

.... p 1 INTRODUCTION ISUAL tracking has emerged as an important component of systems in several application areas including vision based control [1] 2] 3] 4] human computer interfaces [5] 6] 7] surveillance [8] 9] agricultural automation [10] [11], medical imaging [12] 13] and visual reconstruction [14] 15] 16] The central challenge in visual tracking is to determine the image configuration of a target region (or features) of an object as it moves through a camera s field of view. This is done by solving what is known as the ....

....changes from frame to frame. As a result any error in motion estimation between any two frames is subsequently propagated through the entire sequence. Another well established route toward efficient tracking is to detect and track only a sparse collection of features (or contours) 30] [11], 31] 32] 33] As such methods use local detection of areas of high contrast change, they tend to be insensitive to global changes in the intensity and or composition of the incident illumination. However, in many situations persistent, strong edges are sparsely distributed throughout the ....

[Article contains additional citation context not shown here]

 D. Reynard, A. Wildenberg, A. Blake, and J. Marchant, "Learning Dynamics of Complex Motions From Image Sequences," Proc. European Conf. Computer Vision, vol. 1, pp. 357--368, 1996.

....facial displays and expressions) This separation produces an easier tracking problem by requiring a smaller description of object state to be estimated in each frame. This division is often built into face models [6, 28, 34, 47] to simplify model construction or estimation, while Reynard, et al. [40] use this separation to permit learning the variability of motions for a class of objects. Note that there is no guarantee that the shape and motion of some class of objects is separable; this is a simplifying assumption that we make. For human faces, this separation is quite reasonable, and ....

D. Reynard, A. Wildenberg, A. Blake, and J. Marchant. Learning dynamics of complex motions from image sequences. In Proceedings ECCV '96, pages I:357--368, 1996.

....for curve tracking This chapter outlines a probabilistic active contour framework for visual tracking where objects are represented by B spline curves in an image stream. This framework was developed by Blake and a number of collaborators (Curwen, 1993; Blake et al. 1993a; Blake et al. 1995; Reynard et al. 1996; Rowe, 1996; Wildenberg, 1997; Kaucic, 1997; North and Blake, 1998) and is fully set out in (Blake and Isard, 1998) The shape and motion models used are very similar to those adopted by Cootes et al. 1994) and Baumberg and Hogg (1995b; 1995a) While the Condensation algorithm and its ....

....to end with a pause than go straight into a drawing motion. Since the scribbling motion is an oscillator with small spatial extent, a slight variant of Chapter 5. Mixed discrete continuous motion models 89 the standard SDE model was used which allows the means of successive oscillations to differ (Reynard et al. 1996). The concept of a scribble unit is introduced, which is a maximal consecutive subsequence of states fX i g all having discrete label y = 3 informally this is an entire scribbling motion, from start to finish. Each scribble unit is considered to have a fixed mean, but distinct scribbles have ....

Reynard, D., Wildenberg, A., Blake, A., and Marchant, J. (1996). Learning dynamics of complex motions from image sequences. In Proc. 4th European Conf. Computer Vision, 357--368, Cambridge, England.

.... 1 S 11 = 0; and C = B 0 B t 0 is given by C = 1 N 0 Gamma 2 S 22 A 1 S 11 A t 1 A 0 S 00 A t 0 Gamma S 21 A t 1 Gamma S 20 A t 0 A 1 S 10 A t 0 Gamma A 1 S 12 Gamma A 0 S 02 A 0 S 01 A t 1 : This result is extended to allow the learning of x by Reynard et al. [10]; the details are not repeated here. From the point of view of applying EM, the important point is that the log likelihood L is linear in the moments S ij and S i = P N Gamma2 n=1 xn i . This is the Maximisation step; to perform the Expectation step, note that the log likelihood L depends on ....

D. Reynard, A.P. Wildenberg, A. Blake, and J. Marchant. Learning dynamics of complex motions from image sequences. In Proc. 4th European Conf. Computer Vision, pages 357--368, Cambridge, England, Apr 1996.

....the trace of P) and once this has happened, the estimates of these states are fixed; no further observations will alter their values. Although such an approach may at first seem a little curious, there is sometimes good reason to allow the variance of an estimate to collapse to zero. In [8] the mean of a process is estimated; clearly a mean is a single, fixed quantity which does not vary with time. If the estimation process allows accurate evaluation of the mean it is quite correct to use the process covariance to constrain the value of the estimated mean by allowing the estimate ....

D Reynard, A Wildenberg, A Blake, and J A Marchant. Learning dynamics of complex motions from image sequences. In B Buxton and R Cipolla, editors, Computer Vision -- ECCV `96, Lecture Notes in Computer Science. Springer, April 1996.

....and less frequently somewhat longer periods of scribbling y = 3. Also, scribbling motions often start or end with a pause. Since the scribbling motion is an oscillator with small spatial extent, a variant of the standard SDE model was used which allows the mean of the oscillation to vary [11]. Each scribbling motion is considered to have a fixed mean, but distinct scribbles have distinct means. This is encoded by augmenting scribble state samples with an extra vector denoting the mean configuration X scribble = x; 3; x) The translation components of the mean vector x are ....

D. Reynard, A.P. Wildenberg, A. Blake, and J. Marchant. Learning dynamics of complex motions from image sequences. In Proc. 4th European Conf. Computer Vision, 357--368, Cambridge, England, Apr 1996.

.... equations for estimation of the parameters of autoregressive models [Gelb, 1974, Goodwin and Sin, 1984, Ljung, 1987] Suitable adaptations for multidimensional shape spaces are given by [Blake and Isard, 1994, Baumberg and Hogg, 1995b, Blake et al. 1995] with a number of useful extensions in [Reynard et al. 1996]. One example is the scribble in figure 12, learned from the training sequence in figure 9. A more complex example is learning the motions of an actor s face, using the shape space described earlier that covers both rigid and non rigid motion. Figure 13 illustrates how much more accurately ....

....of the learned dynamics are significantly improved when EM is used, especially in the case of more coherent oscillations. An extension of the basic algorithm for classes of objects, dealing independently with motion and with variability of mean shape position over the class, is described in [Reynard et al. 1996]. The same algorithm is also used for modular Phil. Trans. R. Soc. Lond. A (1998) Statistical models of visual shape and motion 13 Figure 13. Trained dynamics for facial motion. Hand built dynamics, exhibited here by random simulation (left) are just good enough, when used in tracking, to gather ....

Reynard, D., Wildenberg, A., Blake, A., and Marchant, J. (1996). Learning dynamics of complex motions from image sequences. In Proc. 4th European Conf. Computer Vision, 357--368, Cambridge, England.

No context found.

D. Reynard, A. Wildenberg, A. Blake, and J. Marchant. Learning dynamics of complex motion from image sequences. ECCV-96, pp. 357--368, Cambridge, UK, 1996. 920

No context found.

Reynard, D., Wildenberg, A. P., Blake, A. & Marchant, J. 1996 Learning dynamics of complex motions from image sequences. In Proc. Eur. Conf. on Computer Vision, Cambridge, April 1996, pp. 357--368.

No context found.

D. Reynard, A. Wildenberg, A. Blake, J. Marchant, Learning dynamics of complex motions from image sequences, Proceedings of European Conference on Computer Vision 1 (1996) 357 -- 368.

No context found.

D. Reynard, A. Wildenberg, A. Blake, and J. A. Marchant. Learning dynamics of complex motions from image sequences. In ECCV, 1996.

No context found.

D. Reynard, A.P. Wildenberg, A. Blake, and J. Marchant. Learning dynamics of complex motions from image sequences. In Proc. 4th European Conf. Computer Vision, pages 357-- 368, Cambridge, England, Apr 1996.

No context found.

D. Reynard, A. Wildenberg, A. Blake, and J. Marchant. Learning dynamics of complex motions from image sequences. ECCV, 1996.

No context found.

D. Reynard, A.P. Wildenberg, A. Blake, and J. Marchant. Learning dynamics of complex motions from image sequences. In Proc. 4th European Conf. Computer Vision, pages 357-- 368, Cambridge, England, Apr 1996.

No context found.

# D. Reynard, A. Wildenberg, A. Blake, and J. Marchant, "Learning Dynamics of Complex Motions From Image Sequences," Proc. European Conf. Computer Vision, vol. 1, pp. 357--368, 1996.

No context found.

D. Reynard, A. Wildenberg, A. Blake, and J. Marchant. Learning dynamics of complex motion from image sequences. ECCV-96, pp. 357-368, Cambridge, UK, 1996.

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