Abstract: A modal heat feature（MHF） is devised to improve the non-ideal retrieval performance of the current heat distribution based non-rigid 3D shape retrieval methods. First, on the basis of the heat kernel related features, a normalized heat distribution function is introduced together with an integral kernel function. Then, a structural heat distribution matrix is constructed. Finally, the largest eigenvalue sequence is employed as the modal heat feature. The proposed feature has many interesting properties, such as intrinsic invariant, discriminative and robust to noise. Experiment results on two public accepted datasets show that the proposed method outperforms all the prior heat distribution based methods on both accuracy and robustness. Besides, compared with the advanced methods, the proposed method has reached high retrieval performance and robustness.