This paper presents a method for key-frame extraction from 3D video using visual saliency to weight the 3D content according to a user attention model. Key-frames are found in temporal segments of arbitrary length (i.e., 3D scenes) using a dynamic programming algorithm which minimises the dissimilarity between the reconstructed and the original temporal segment. The dissimilarity measure is based on a combination of frame difference and visual relevance estimated through visual saliency maps. These maps result from attention modeling, taking into account spatial, temporal and depth features of the 3D video content. The results, evaluated using the Shot Reconstruction Degree and the Fidelity measure, show that the proposed method outperforms those obtained from uniform sampling and attention curve methods. This method may be useful for fast browsing of 3D video repositories.