Research On Camera Self-Calibration For3D Reconstruction With Varying Intrinsic Parameters

Abstract:3D Reconstruction is a hot research field in computer vision. Traditional methods of3D Reconstruction require calibrated image sequence which limits the data resource. The rise of the internet and the advent of the photo sharing community bring huge amounts of pictures taken all over the world. How to reconstruct from those internet images and other uncalibrated images? Self-calibration based on scene information provides a solution. This paper concerns about self-calibration methods with uncalibrated or partial calibrated image sequence based on the dual absolute quadric is fixed under Euclidean motion of camera. According to the application above, two problems as follow are solved:(1) Extracting focal length to each view with known other intrinsic parameters;(2) Extracting intrinsic parameters to each view with image sequence from internet which contains an identity scene. Meanwhile, fewer constraints assumption would be better.According to the problems above, two algorithms have been proposed:(1) A varying focal length self-calibration method with known other intrinsic parameters is proposed. Firstly, a projective reconstruction is computed from point correspondences across the image set. Then, the dual absolute quadric is achieved form constraints. The dual absolute quadric matrix is rectified globally by bundle adjustment. Once the dual absolute quadric is identified, focal length is determined. Experiments demonstrate that the proposed method can extract accurate focal length to each view.(2) A varying intrinsic parameters self-calibration method based on internet image sequence is proposed. Firstly, a projective reconstruction based on trifocal is computed. Then extract focal length, manufacturer and model number to partial images by EXIF tags. Classify all images into different classes by manufacturer and model number, pre-calibration is applied to each class. Bundle adjustment is used to achieve global optimization. Experiments demonstrate that the proposed method can accurately calibrate images sequence from internet.