Citation Context

...del (FHMM) where observations have been marginalized out and is quite tractable. Exact inference requires triangulating the graph (figure 2-(e)) to make explicit additional probabilistic dependencies =-=[11]-=-.The maximum clique size is now 3, making inference O(k 2 cN) (where N is the number of total frames). Furthermore, the triangulation affords us the opportunity to model additional dependencies betwee...

Citation Context

...ynthesized sequence to the video. Previous work is extensive, as classifying human motions from some input is a matter of obvious importance. Space does not allow a full review of the literature; see =-=[1, 5, 4, 9, 13]-=-. Because people do not change in appearance from frame to frame, a practical strategy is to cluster an appearance model for each possible person over the sequence, and then use these models to drive ...

Citation Context

...ynthesized sequence to the video. Previous work is extensive, as classifying human motions from some input is a matter of obvious importance. Space does not allow a full review of the literature; see =-=[1, 5, 4, 9, 13]-=-. Because people do not change in appearance from frame to frame, a practical strategy is to cluster an appearance model for each possible person over the sequence, and then use these models to drive ...

Citation Context

...ynthesized sequence to the video. Previous work is extensive, as classifying human motions from some input is a matter of obvious importance. Space does not allow a full review of the literature; see =-=[1, 5, 4, 9, 13]-=-. Because people do not change in appearance from frame to frame, a practical strategy is to cluster an appearance model for each possible person over the sequence, and then use these models to drive ...

Citation Context

... to k = 300 frames by clustering with k-means and retaining only the cluster medoids. Our metric is based on differences between 3D key points, velocities, and accelerations for the 3D motion frames (=-=[2]-=- found this metric sufficient to ensure smooth motion synthesis). The motion capture data are represented atT m M 1 t 1 M m 1 T 1 1 M T 1 t Undirected Variables Directed model model Factorial HMM Tri...

Citation Context

...h individual is differently dressed, so that the number of individuals is the number of distinct appearance models. Detecting the learned appearance model in the sequence of frames is straightforward =-=[8]-=-. 4. 3D Motion Synthesis Once the 2D configuration of actors has been identified, we need to synthesize a sequence of 3D configurations matching the 2D reports. Maintaining a degree of smoothness — i....

Citation Context

...nce (section 5), accepting them as annotations for the underlying video sequence. 2. Obtaining Annotated Data We have annotated a body of motion data with an annotation system, described in detail in =-=[3]-=-; we repeat some information here for the convenience of the reader. There is no reason to believe that a canonical annotation vocabulary is available for everyday motion, meaning that the system of a...

Citation Context

...ssified; except in the case of specific activities, there is no known natural set of categories. Special cases include ballet and aerobic moves, which have a clearly established categorical structure =-=[5, 6]-=-. In our opinion, it is difficult to establish a canonical set of human motion categories, and more practical to produce a system that allows easy revision of the categories (section 2). Figure 1 show...

Citation Context

...rance model for each possible person over the sequence, and then use these models to drive detection. This yields a tracker that is capable of meeting all our criteria, described in greater detail in =-=[14]-=-; we used the tracker of that paper. Leventon and Freeman show that tracks can be significantly improved by comparison with human motion [12]. Describing motion is subtle, because we require a set of ...

Citation Context

...orso and long arms; the tendency of ratios of body segment lengths to vary from individual to individual and with age is a known, but not well understood, source of trouble in studies of human motion =-=[10]-=-. Our motion capture database is too large for us to use every frame in the matching process. Furthermore, many motion fragments are similar — there is an awful lot of running — so we vector quantize ...

Citation Context

...ynthesized sequence to the video. Previous work is extensive, as classifying human motions from some input is a matter of obvious importance. Space does not allow a full review of the literature; see =-=[1, 5, 4, 9, 13]-=-. Because people do not change in appearance from frame to frame, a practical strategy is to cluster an appearance model for each possible person over the sequence, and then use these models to drive ...

Citation Context

...f meeting all our criteria, described in greater detail in [14]; we used the tracker of that paper. Leventon and Freeman show that tracks can be significantly improved by comparison with human motion =-=[12]-=-. Describing motion is subtle, because we require a set of categories into which the motion can be classified; except in the case of specific activities, there is no known natural set of categories. S...

Citation Context

...ince the motion is sampled in time, each joint has a discrete 3D trajectory in space for the second of motion centered at the frame. In our implementation, we used a public domain SVM library (libsvm =-=[7]-=-). The out of margin cost for the SVM is kept high to force a good fit within the capabilities of the basis function approximation. Our reference collection consists of a total of 7 minutes of motion ...