Human Motion De-noising via Greedy Kernel Principal Component Analysis Filtering

Kernel principal component analysis (KPCA) has been shown to be a powerful non-linear de-noising technique. A disadvantage of KPCA, however, is that the storage of the kernel matrix grows quadratically, and the evaluation cost grows linearly with the number of exemplars. The size of the training set composing of these exemplars is therefore vital in any real system incorporating KPCA. Given long human motion sequences, we show how the greedy KPCA algorithm can be applied to filter exemplar poses to build a reduced training set that optimally describes the entire sequence. We compare motion de-noising between standard KPCA using all poses in the original sequence as training exemplars and de-noising using the reduced set filtered by the greedy algorithm. We show how both have superior de-noising qualities over PCA, whilst Greedy KPCA results in lower evaluation cost due to the reduced training set