To render this inversion feasible, we applied a sparse approximation method that approximates the marginal distribution based on a low number of inducing points in the hidden spaces. The model was trained using six motioncaptured example trajectories for each of the two basic human and monkey expressions, sampled with time steps.They had to discriminate between animations with original motion capture data and ones generated with movements that were generated by the morphing algorithm. The movements generated with the algorithm approximated the prototype movements. In order to induce some variability, we used three different motion capture trials of each of the original human and monkey expressions, and their approximations based on the morphing model.The compared stimulus pairs were presented sequentially, and motions were presented in a blockrandomized order times, in separate blocks for the two avatar types.To verify that participants can pick up artifacts in the animations, we added a further condition in which instead of movements generated by the morphing algorithm we used control movements, which were generated by reversing the temporal order of short fourframe segments in the original motioncaptured movements.Animations with these control movements also had to be distinguished from ones with the original motion capture data.The accuracy of the detection of original motion capture data, as opposed to the generated one, was. for the human avatar, both values were significantly lower. This implies that the animations using motion capture data were judged even less frequently as original trajectories than the animations generated with our motion synthesis algorithm.The morphing algorithm thus does not degrade the perceived naturalness of the motion.According to this criterion, for almost all stimulus types, model four was the most compact model for the front view stimuli. Only for the rotated views of the avatars, however, we found a slight significant increase of the prediction accuracy. For this reason, we decided to use model four as the basis for our further analyses of all classification data in the main experiment.In order to control for this variable, we created a second stimulus set for which the amount of lowlevel information was balanced.The stimuli for this experiment were motion morphs between each of the four prototype expressions and a neutral expression.The original expression entered the motion morph with a <a href="https://www.ncbi.nlm.nih.gov/pubmed/21507972">search Pomalidomide-PEG3-C2-NH2</a> weight of l, and the neutral expression with a weight of l, where the morphing weight was adjusted to obtain the same amount of lowlevel information in all adjusted prototypes.We found that the measures DF and MF, <a href="https://www.targetmol.com/compound/Pomalidomide-PEG3-C2-NH2">buy Pomalidomide-PEG3-C2-NH2</a> computed from the twodimensional and threedimensional mesh coordinates, and the control points were very highly correlated. The mesh pointbased measures were monotonically increasing functions of the morphing level l.This was not the case for the quantities computed from the control point trajectories, due to which we discarded the measures derived from the control point trajectories from the balancing of the stimuli.Because of the high correlation between the measures computed from the twoand threedimensional meshpoint trajectories, and the higher similarity of the twodimensional trajectories with image motion, we kept only the measures computed from the twodimensional meshpoint trajectories for the further analysis.
