Abstract: The virtual binocular vision sensing system was designed based on single camera through the refraction principle of biprism, the accuracy of which reached to sub-pixel. The intrinsic and external parameters of both virtual cameras and relative position were calculated, and the image pairs were rectified. Then, adaptive Census transform (ACT) stereo matching algorithm was optimized for calculating the parallax of the rectified image pairs with sub-pixel accuracy. The three-dimensional appearance of a standardized cylinder was reconstructed through the triangulation method. The results showed that the width error was less than 3.17% and the height error had a maximum of 5.83%. The image pairs of molten pool in GMA additive manufacturing were taken as an example, while the parallax with sub-pixel accuracy of the image pairs was calculated based on the optimized adaptive Census transform stereo matching algorithm. Finally, the three-dimensional appearance of molten pool was obtained via the triangulation approach.