Specimen Pan Angle Estimation from Orthographic Multi-View Images

Image-based 3D reconstruction uses the image information, more precisely the depth cue, in one or more images to model an object in three dimensions. A wide variety of approaches and combinations with respect to the experimental setup for acquiring a data set, the depth cue and the processing of the image data have already been extensively evaluated. In general, image-based 3D reconstruction approaches require calibration of the camera(s) to properly process the image information. This work investigates an approach to image-based 3D reconstruction that realizes precise three-dimensional modeling without calibration of the camera(s). The underlying setup for acquiring the data sets consists of a static camera and a turntable to which the object is attached. Correction of the image information based on the pan angle of the object with respect to the rotation axis of the turntable implements the appropriate association of the image data without calibration of the camera. In the context of 3D reconstruction, the use of orthographic images allows direct determination of the dimensions of the object. Nevertheless, the distortion of the measurements due to the inclination of the object orthogonal to the image plane is a problem. This inclination of the object with respect to the image plane and consequently the distorted measurements can be corrected based on the assumptions made about the experimental setup for image acquisition and the object. An artificial data set, which shares the characteristics of the data set of a real specimen, is used for the evaluation. In addition, the application to the data set of a real specimen is analyzed in detail. The results of the work illustrate that the assumptions made and the approaches chosen realize a very accurate digital reconstruction for both the artificial data set and the data set of a real specimen. Optimizing the object's pan angles in the global context w.r.t. a data set minimizes the influence of random errors. The precision of the 3D model increases with the number of views of an object.