for both the humanoid and CG character. Technique
EPH yielded extreme motion most likely because the
implementation, based on linear interpolation of the
lambda parameters and an exponential force gener-
ation equation, did not include a velocity damping
term so that motions tended to bo unstable.
A final note is that wo attempted to find an average
solution which fairly represented the data. However,
duo to the distinctive nature of each of the outliers in
the 4 different conditions, attempts at this revealed
that the overall solution lost important characteris-
tics of the individual solutions.
Discussion. Overall the results indicate that six
movements account for most of the variability in the
data. Those are 1) MT for humanoid movement (a),
2) MT for CG movement (a), 3) EPH for humanoid
movement (b), 4) EPH for CG movement (b), 5)
MTP for humanoid movement (b) and 6) MTPVT
for humanoid movement (b). Informal examination
of the movies reveal that MT for movement (a) is
unnatural duo to several arm swings inititating the
movement from start to endpoint; EPH for move-
ment (b) is a very indirect path and MTP and MT-
PVT for movement (b) appear awkward on the hu-
manoid duo to apparent difficulty of the humanoid to
obtain one of the intermediate postures. This awk-
wardness wasn’t apparent on the CG character move-
ment. For the first 4 cases the hand path was very
different from the other movements and thus the dis-
similarity results are consistent with oven informal
observation of the sot of movements. However, for
the other two cases it is not obvious what physical
properties of the movements are making them appear
distinct.
f,.2 Experiment 2 - Pairwise ratings of the
CG character with outliers removed
In Experiment 2 wo examined ratings of dissimilar-
ity to CG character movements. Wo used the same
two movements as used in Experiment 1 however for
production techniques wo excluded the EPH and MT
technique since they dominated the MDS solution,
precluding the possibility of other factors explaining
participants’ responses. Wo chose the CG charac-
ter rather than the humanoid since, as indicated by
MTP and MTPVT for movement (b), the humanoid
appeared to demonstrate additional properties in the
movement performed. Altliougli just what additional
properties the humanoid adds to the movement is of
interest it is beyond the scope of the current work.
Design and methods. A total of 7 volunteers
participated in the study. Each participant made
pairwise dissimilarity judgements on the two sots of
78 motion pairs generated form the sot of 13 versions
of movement (a) and movement (b). Each sot of 78
movements was presented together in a block and a
small rest occurred between blocks.
Results. The results of each individual partici-
pant was averaged together to form an estimate for
each movement pair and the two resulting dissimilar-
ity matrices wore input to an INDSCAL algorithm
to find a two dimensional solution. The solution
obtained had a stress of 0.15 and r-squarod value
of 0.94 with dimension 1 having an overall impor-
tance of 0.72 and dimension 2 having an importance
of 0.22. The results of this can bo soon in Figure
2. An additional piece of information provided by
the INDSCAL algorithm that is relevant to the cur-
rent experiment is the weighting of dimensions for
the two matrices. This revealed that movement (a)
had a dimension 1 weight of 0.93 and a dimension
2 a weight of 0.25, movement (b) had a dimension 1
weight of 0.72 and a dimension 2 weight of 0.62. Indi-
cating that the fast movement (movement (a)) was
evaluated primarily along a single dimension while
the slow movement (movement (b)) was evaluated
equally along two dimensions.
Figure 2: Results of the INDSCAL algorithm for the fits
to participants pairwise dissimilarity judgments
Overall the structure revealed in Figure 2 indicates
that the MTP and MTPVT production techniques
wore outliers on both dimension 1 and dimension 2.
In addition the human motion capture movements
appeared to not bo grouped with any of the other
production techniques. In the Discussion wo will
examine the implications of this as well as consider
what physical properties might undorly the two di-
mensions.
Discussion. MDS provides a useful technique for
finding a compact description of a largo sot of mea-
surements - reducing a largo number of pairwise com-
parisons to the positions of the stimuli in a low di-
mensional space. However, it doos not automatically
provide an interpretation of the dimensions. Some
obvious candidates to consider as explanatory vari-