The quantitative analysis of self-calibration based on rotating cameras

Self-calibration based on rotating cameras is a classical technique for recovering internal parameters which ensures that the camera motion is pure rotation and does not permit translation. In practice, however, it is hard to exactly keep the camera not to translate. Previous studies on this issue are mainly qualitative and less operable for practice applications. In this paper, we experimentally develop a quantitative analysis of the influences of the camera translation, image noises and the number of image on the accuracy and stability of the method. Many quantitative results are obtained, for example, when the ratio between the translation of camera and the depth of space points is 1/200, and the image noise is 2 pixels, we need at least 6 images for stable result with higher accuracy. All these results can give users a more operable guideline in real applications.