Abstract: In this paper, a novel approach is presented for motion retrieval based on double-reference index that reduces the number of costly distance computations for similarity measure. In order to retrieve motion data accurately, the features about joint positions, angles and velocities are extracted to represent motion data. Since these original features of motion clips lie in high-dimensional space and on a high-dimensional manifold, which is highly contorted, the Isomap nonlinear dimension reduction is used to map them into low-dimensional manifold. However, geo-distance of Isomap is only defined on training sets and raw Isomap cannot map new samples to the embedding subspace because it requires a whole set of points in the database to calculate geo-distance. For handling out-of-samples motion data, Isomap is generalized based on the estimation of underlying eigenfunctions. Then the methold of double-reference index(DRI) is proposed based on selecting a small set of representative motion clips in the database. So it can get candidate set by abandoning most unrelated motion clips to reduce the number of costly similarity measure significantly. Finally the automatic method is tested on a large collection of motion capture clips with a large variety of actions. Experiment results show that the proposed methods are effective for motion data retrieval in large-scale database.