The presented study tests which type of stereoscopic image, natural or artificial, is more adapted to perform efficient and reliable calibration in order to track the gaze of observers in 3D space using classical 2D eye tracker. We measured the horizontal disparities, i.e. the difference between the x coordinates of the two eyes obtained using a 2D eye tracker. This disparity was recorded for each observer and for several target positions he had to fixate. Target positions were equally distributed in the 3D space, some on the screen (with a null disparity), some behind the screen (uncrossed disparity) and others in front of the screen (crossed disparity). We tested different regression models (linear and non linear) to explain either the true disparity or the depth with the measured disparity. Models were tested and compared on their prediction error for new targets at new positions. First of all, we found that we obtained more reliable disparities measures when using natural stereoscopic images rather than artificial. Second, we found that overall a non-linear model was more efficient. Finally, we discuss the fact that our results were observer dependent, with variability’s between the observer’s behavior when looking at 3D stimuli. Because of this variability, we proposed to compute observer specific model to accurately predict their gaze position when exploring 3D stimuli.