Structure-oriented filtering is based on anisotropic diffusion equations, which controls the direction and speed of filtering through constructing diffusion tensor by structure tensor. According to the discussion on the impact of different types of diffusion tensor on the filtering effect of diffusion equations, robust structure tensor was obtained by the convolution between Gaussian kernel and first-order partial differential operator, which introduced the method of calculating 3D structure-oriented filtering. The results prove that the method is effective to improve signal-to-noise ratio of seismic data, as well as correctly extract or analyze volumetric curvature attributes.