In this paper, a novel tensor dimensionality reduction approach for 2D+3D facial expression recognition via orthogonal Tucker decomposition using factor priors (OTDFPFER) is proposed. First, a 4D tensor model is built by stacking six effective features from 2D images and 3D face data to maintain their structure information and correlations, which simultaneously avoids the dimension disaster and small samples. Second, from the perspective of geometry, the real-world high dimensional tensors often lie in a low dimensional subspace, Tucker decomposition as a powerful technique is used to extract the useful information from the generated 4D tensor, aiming to achieve a set of core tensors with smaller sizes and a set of factor matrices for projecting into the generated 4D tensor. Third, during the 4D tensor modeling process, information will be partially missed, and high similarities among samples will emerge. Based on the tensor orthogonal Tucker decomposition, the structured sparsity of the involved core tensor, and a graph regularization term via the graph Laplacian matrix together with the $4$th factor matrix are employed to better characterize these similarities. Meanwhile, a tensor completion (TC) framework is embedded to recover the missing information. An alternating direction method coupled with the majorization-minimization scheme is designed to solve the resulting tensor completion problem. Meanwhile a cutting strategy is utilized to strength the interactions among the core tensor and factor matrices and to shorten the convergency process. The numerical experiments are conducted on the BU-3DFE and Bosphorus databases with promising recognition accuracies.