Abstract: The credibility and precision of 3D reconstruction are sensitive to registration errors and rectification errors. To tackle this problem, we propose a high precision rectification-free narrow base-line 3D reconstruction algorithm. The high accuracy parallax is acquired by combining the phase-correlation and curve fitting method. A peak obvious matrix is obtained by inverse Fourier transform of normalized cross-power spectrum which is generated from Fourier transformations of two images. Measurements of disparity at sub-pixel level is obtained by curve fitting data near that peak. The base-line's projection on the imaging plane (called main-direction for short) is found by computing and normalizing the sum of all points' parallaxes. The 3D Euclidean reconstruction is obtained by projecting all points' parallaxes onto the main-direction because the depth value of each point is proportional to its projection. A large number of experiments on all kinds of scenes prove that the proposed algorithm can attain excellent reconstruction performance for stereo images with narrow baseline.