Comparing Self-calibration Methods for Static Cameras

Extended Abstract Many methods have been developed in the last few years to self-calibrate cameras , but few works have addressed a comparison of such methods to provide the user with hints on the suitability of certain algorithms under particular circumstances. The few of these works that analyze self-calibration methods have concentrated on the study of the influence of some factors [1] [2] or the identification of critical movements [3], but there are not any studies analyzing the accuracy and stability of these methods. This work presents a comparative analysis of four methods of self-calibration for cameras which only rotate (without translation movement): McLauchlan's method [4] that assumes that the camera setup is fixed along the sequence; and Agapito's methods (linear [5] and iterative [6] algorithms) and Seo's method [7] that allow the variation of optic center and scale factors. The experiments discussed in this paper have focused on characterizing the accuracy in the point reconstruction (global error) and the stability and accuracy of the estimation of the internal camera parameters. These experiments were performed with both a real camera and a simulator. When only the extrinsic parameters vary in a sequence used to calibrate a camera the estimated values for the intrinsic parameters in each image of this sequence should be the same. The variability in the estimation of the intrinsic parameters along the sequence was measured to know the stability of the methods. The experiments carried out showed that, despite the camera configuration was kept constant along the sequence, there were variations across all methods by intrinsic parameters. For example, this variability represented 4% of the ground truth value in the optic center when simulated images was used and 3% when real images were analyzed. Additionally, the estimated values of intrinsic parameters obtained by the methods had differences between them, especially in the scale factors (4% using simulated data and 2.5% using real data). In the simulator, the accuracy of each parameter could be analyzed. It was observed that the versions of Agapito's method obtained values closer to ground