The proposed framework to compute Eigen Joints is demonstrated in Fig. 2. We employ 3D position differences of joints to characterize action information including posture feature, motion feature , and offset feature in each frame. We then concatenate the three features channels as ݂Nmm, According to different experimental settings in Section 5.1, two normalization schemes are introduced to obtain . In the end, PCA is applied to D to compute EigenJoints.
To characterize the static posture information of current frame, we compute pair-wise joints differences within
the current frame:
To represent the offset feature or the overall dynamics of the current frame-ܿ with respect to the initial frame, we calculate the pair-wise joints differences between frame and frame The initial frame tends to approximate the neutral posture. The combination of the three feature channels forms the preliminary feature representation for each frame