A geometric approach to parameter estimation from tomographic data

This paper presents a method for fitting surface models of solid objects directly to 2D projections without the need for detecting intermediate features. The method relies on first principles and follows from an analytical formulation of the change in the projected image with respect to the object parameters. It incorporates a gradient descent with an inverse-Hessian minimization procedure to efficiently find local minima in an image-based misfit function. The same formulation applies to system parameters, and thus one can simultaneously fit a surface model and estimate the distortion parameters of the sensor. Results are shown for simulated data and real images from a C-arm fluoroscopy device.