Figure 1. Tracking a Flexing and Translating Finger

Figure 2. Condensation tracker (above) vs. subset tracker (below). mode, namely no motion at all, was tracked; the condensation tracker is unable to deal with this, while the subset tracker, which does not rely on dynamics, is successful. frame 6 frame 13 Figure 3. Condensation tracker (above) vs. subset tracker (below). There are two principal directions for future research. First, the development of an algorithm for learning a mul-tidimensional manifold would be of benefit. In the current experiments, a one-dimensional manifold was learned in invariant space; however, it is quite likely that the true man-ifold was of higher dimension. (In this scenario, the one-dimensional manifold is simply a subset of the higher dimensional manifold.) Success in this area would also allow for more efficient implementation of the algorithm. Second, the algorithm may be extended to the task of object localiza-tion, in which an object is to be located within a single still image, rather than a video-sequence. Edge-search can no longer be initiated at the previous frame's contour estimate, as there is no such estimate; thus, in principle any edge-point in the image may be potentially part of the relevant curve. In other words, for each , the set is very much larger than in the case of tracking. The ability to search through the resulting huge space of observed curves ¡ relies on the © term in the complexity, as discussed in section 4. References [1] A. Blake and M. Isard. Condensation-conditional density propagation for visual tracking. Int.