In this paper, we propose a new method to automatically extract keyframes from animation sequences. Our method can be applied equally to both skeletal and mesh animations. It uses animation saliency computed on the original data to help select the group of keyframes that can reconstruct the input animation with less perception error. For computational efficiency, we perform nonlinear dimension reduction using locally linear embedding and then carry out the optimal search in much lower-dimensional space. With this approach, reconstruction of the animation from the extracted keyframes shows much better results as compared with earlier approaches. Copyright © 2012 John Wiley & Sons, Ltd.