A monocular stereo vision-based geometric error identification method for machine tools rotary axes is proposed to consider both cost and accuracy of vision measurement. To accurately describe the full-range motion information of the rotary axes, an error amplification-based feature target with the precise low-reflection target balls in circumferential uniform distribution is designed to guarantee the high-uniformity, high-contrast, and high signal-to-noise ratio (SNR) imaging. A simulation analysis method for monocular stereo imaging optical path is proposed to quantitatively analyze and select system structure parameters. With the Zhang’s calibration results, the three-dimensional trajectories of rotary axes can be reconstructed from the acquired sequence images. The accuracy verification experiments and error identification results show that the root mean square error (RMSE) of 0.049 mm is less than 1/3 of the mean measurement error, satisfying the test requirements.