Reconstructing and registering 3D surface from single intensity image
MetadataShow full item record
Reconstructing 3D surface from of an intensity image is a fundamental problem in the computer-vision community. Humans have a remarkable ability to perceive shape of an object when looking at a 2D intensity image, but enabling computers to do so is still a very challenging task. Advances in this area can help us solve problem of building panoramic 3D view by registering multiple surfaces. It has many useful applications such as medical surgery using CT/MRI scans, restoring ancient artifacts using single image, path/arm guiding algorithms for robots, etc. Most existing methods in this area belong to the category of Shape from Shading. Such methods have shown limited success recovering simple shapes, such as a vase. However, due to inherent ambiguity associated with relating intensity to depth, existing methods have not been able to demonstrate similar success for complex surfaces. Improved methods use additional information, such as stereo images like Structure from Motion, or multiple light sources like Photometric Stereo. Such methods have two broad issues. Firstly, inherent problem from Lambertian reflectance model is not solved and therefore, secondly, they could not provide insights regarding the human ability to perceive depth from 2D images. In this thesis, we are trying to solve the problem by exploiting relationship between surface gradient and corresponding intensity value at the exact position. The relationship is formulated as a novel illumination model and the model is at the core of our research. Using the illumination model, surface reconstruction problem is posed as a task of integrating step-wise absolute value of difference of depth (think of gradient). To solve the inherent ambiguity (even human being could not decide), we tackle the relative signs by using reference depth as a prior knowledge. Then the formula is simplified and approximated for handing different materials and lighting conditions. Based on the reconstructed surfaces, we can also handle the problem of building a panoramic 3D surface with intensity images of same object from different views. The core step here is surface registration. We developed methods that use Riemannian Geometry and Machine Learning techniques for rigid or non-rigid surface registration.