Optical polarization is three-dimensional (3-D). Its complete information is described by the nine-component generalized Stokes vector (GSV). However, existing Stokes polarimetry and its design theory are primarily based on the paraxial four-component Stokes vector and 4 × 4 Mueller matrices. In this Letter, we introduce a novel concept of generalized Stokes polarimetry (GSP), which can reconstruct nine generalized Stokes parameters through a series of non-paraxial polarized modulations and intensity projections. The reconstruction theory of GSP is based on the 9 × 9 generalized Mueller matrix (GMM) calculator we reported previously. In addition, to optimize the 9 × 9 analysis matrix of GSP, we developed an optimization algorithm combined Monte Carlo and gradient descent (GD) methods, finding the optimal configuration with CN = 3.7261 and EWV = 1.2405. The simulated results of noise sources and GSV reconstruction verified the significant improvement in accuracy and stability of optimized configuration.