Information Thoretic Criterion AIC and Camera Calibration Slide
Information Thoretic Criterion (AIC) and Camera Calibration Slide
• G. Wei, K. Arbter, G. Hirzinger, “Active self-calibration of robotic eyes and handeye relationships with model identification, ” IEEE Transaction Robotics and Automation, vol. 14(1), pp. 158 -166, Feb. 1998. • Statistical inference based on Fisher’s distribution. • Select the best model to describe a problem while minimizing the number of parameters. Increase the reliability of the estimated parameters. • Avoid over-parameterization which can increase variance. • Three hypotheses are tested based on distortion parameters complexity (total of two tangential coefficients and one radial coefficient): – Without radial and tangential distortion. – Without tangential distortion but with radial distortion. – With both radial and tangential distortion. 10/3/2020 Slide 2/29
• M. T. El-Melegy and A. A. Farag, “Nonmetric lens distortion calibration: closedform solutions, robust estimation and model selection, ” in Ninth IEEE Int. Conf. on Computer Vision, 2003. vol. 1, pp. 554 -559, 13 -16 Oct. 2003. • Geometric inference. • Criterion defined in terms of the fitness of the data to the model and the complexity of the model. • Complexity of the model is dependent on the number of lines used for distortion calibration and the number of points used. • Three models are tested with AIC and MDL on synthetic data: – Without distortion. – One radial distortion coefficient. – Two radial distortion coefficients. 10/3/2020 Slide 3/29
• M, Ahmed and A. Farag, “Nonmetric Calibration of Camera Lens Distortion: Differential Methods and Robust Estimation, ” in IEEE Transactions on Image Processing, vol. 14, Iss. 8, pp. 1215 -1230, August 2005. • In the conclusion the authors plans are to incorporate statistical model selection. 10/3/2020 Slide 4/29
• [Broaddus 05] C. Broaddus, “Universal Geometric Camera Calibration with Statistical Model Selection, ” Masters Thesis, Department of Electrical Engineering, The University of Tennessee, Knoxville, TN 2005. • Show results on real wide angle data and not just synthetic data. • Implement AIC for five model selection criterions (AIC MDL BIC SSD CAIC) others show it for only AIC and MDL. • AIC model is penalized only for the number of distortion coefficients. AIC model is not penalized for features found in the image (number of images and number of points). 10/3/2020 Slide 5/29
- Slides: 5