The current state-of-the-art deep learning vision networks commonly employs synthetic approaches for data augmentation when confronted with scenes requiring motion blur. However, existing blur synthesis methods often fall short in accurately simulating motion blur as observed in real-world scenarios, consequently hindering the generalization capability of trained deep visual networks to real-world applications, a phenomenon known as domain shift effects [1]. To address this problem, we propose a novel non-uniform motion blur synthesis method for data augmentation. First, we randomly generate the camera’s motion trajectory using a more general six-degree-of-freedom (6-DOF) camera motion model, and then map this trajectory to pixel-level blur kernels. To efficiently perform spatially varying convolution, we employ non-negative matrix factorization (NMF) to decompose those blur kernels into a set of kernel basis and their corresponding mixing coefficients. This enables parallel execution of spatial variation convolutions, thereby significantly improving the efficiency of blur synthesis. Our experiments demonstrate consistently superior results of the proposed method on publicly available real datasets RealBlur, as well as synthetic datasets GoPro and REDS.