On the neural computation of the scale factor in perspective transformation camera model

The perspective transformation based on pinhole camera geometry is widely used in 3D computer vision. The image coordinates projected by the camera model for a given 3D point are in a 3-tuple, (su, sv, s), where s is a scale factor. The inhomogeneous image coordinates u and v can then be determined by simply dividing the first two elements with the scale factor. Although it is easy to compute a scale factor using a (3×4) camera matrix, the computed s does not correspond with the real physical value of the model; the z coordinate of the projected 3D point in the camera-centered coordinate system. In this paper, we propose a unique neural network structure and its learning algorithm to compute the scale factor of a 3D point. Since the proposed method can estimate the scale factor as the real z coordinate, further vision processing such as camera calibration can be performed efficiently using the value. In our computer simulation, the proposed neural network operated well with proving its validity.