Abstract: Aiming at the issue of low accuracy of single feature model in traditional 3 D shape recognition methods, this paper proposes an adaptive sparse coding fusion algorithm for non-rigid 3 D shape classification. This method first extracts the low-level features of average geodesic distance(AGD), heat kernel signature(HKS), and shape diameter function(SDF) to construct complementary shape descriptors. Secondly, the AGD-BoF, HKS-BoF, and SDF-BoF are generated respectively based on the bag-of-feature(BoF) model. And then we build a weighted matrix via random training. Finally, we adaptively create a fusion between the optimized sparse coding and weighted matrix which is used to achieve effective classification by using the Softmax classifier. A series experiments on two non-rigid databases SHREC10 and SHREC11 have shown that our proposed algorithm has better classification accuracy and stronger robustness.