(a) Undistorted predetermined image. (b) Anamorphic projected image corresponding to viewer position ( 10 , 0 , 2 ) . To see the anamorphic image undeformed, look at it from the right side with the face just above the plane of the paper. 

Context in source publication

Context 1

... A and B are camera-specific constants that depend on the field of view of the camera lens and the resolution of the camera picture sensor and x is the estimated IPD in pixels. The camera that we used throughout all of our experiments reported in this article is a Logitech WebCam Pro 9000 with a horizontal field of view 63 . 1 ◦ , a vertical field of view 49 . 4 ◦ and an image resolution of 1600 × 1200 pixels. We experimentally set A to 21528.8 and B to − 7 . 78 using standard numerical packages. The accuracy of the distance estimation function F ˆ dist was evaluated by comparing the computed distances with the actual measured distances between the observer and the camera using a tape meter at several distances between 0.3 and 10 m. Two participants were asked to stand in front of the camera and consecutively move away from it. After each move, the distance between the camera and a participant increased by 10 cm and a picture of the scene was taken. Fig. 4 presents experimentally measured distances and distances computed using F ˆ dist based on IPD. If the observer of the image needs to keep eye contact with the person on the displayed anamorphic image, it is important that the generated anamorphic image resolves into the correct direction and that the directional error is smaller than 5 ◦ (Stokes, 1969). For the camera configuration that we used, if the face of the observer is 0.5 m away from the camera, a 4-cm shift in lateral direction is already enough to reach the 5 ◦ limit for losing eye contact.At a distance of 3 m the corresponding shift grows to 26 cm and at 10 m the permissible lateral error of position can be up to 87 cm before we lose the eye contact. The error in direction z influences only the scale of the displayed image which does not have any effect on the possibility of establishing eye contact. Fortunately, while the error of position estimation grows with the distance from the camera, the need for lateral accuracy decreases, both due to the same geometrical consideration about the diverging spatial angles and image quantization. A change of position of the observer closer to the camera requires larger anamorphic deformations than when the same change happens farther away. After estimating the distance between the observer and the camera as well as angles ω x and ω y , we have all the information needed to estimate the observer’s position in 3D space. The described method operates in real time and achieves sufficient accuracy for the given problem. Presented concepts were implemented in a standalone application. System uses face detection to determine the approximate 3D position of the observer and uses it to calculate the anamorphic deformation of the projected image thus making it look undistorted to the observer regardless of his position (see Fig. 5). The face detection part uses OpenCV library (Baggio et al. , 2012), while the image is rendered with the help of Mesa library (Surhone et al. , 2010). The application runs on a machine with Intel Core2 Quad Q9000, 2.00 GHz with 4 GB of RAM using NVidia Quadro FX 2700M graphic card. The application is suitable for real-time ...