Introduction: Most of the computer vision techniques used to analyze the scoliosis deformities are based on studying the internal deformation of the spine. The imaging techniques used in these algorithms are mostly based on MRI and Xray imaging. Repetitive exposure to these imaging techniques could yield severe side effects.

Recent approaches for assessing spinal deformities study the exterior surface topography of the torso. Photogrammetry can be used in the measurement of the patient's torso and in the surface reconstruction of the human back.

Objective: This paper contributes to this line of research by developing an intensity based image registration method to align smooth featureless images of the human back. The output of the registration process is then used to reconstruct the shape of the human back.

Materials and methods: A modified coarse-to-fine scheme is adopted in order to contend with larger deformations while still capturing small deformations. A steerable pyramid is built for both the source and target images. The calculated field is accumulated with the field at the next level of the pyramid. A new field is computed at new level and the process is repeated through each level. To register stereo images, epipolar constraints are employed in the registration process. The proposed method put into consideration the change in object appearance due to lighting conditions.

Results: Images for volunteers at the University of Alberta are used in our experiments. Additionally, three dimensional objects with known dimensions are used to measure the accuracy of the stereo reconstruction process.

Conclusions: In this paper, new framework for reconstructing the surface of the human back for scoliotic patients is proposed and tested. Experimental results show the ability of optical flow registration algorithms to generate accurate disparities that could be used in the 3D reconstruction problem.