Non-rigid image registration is one of the most popular problems in computer vision. Many different approaches have been used to recover the transformation between two images. Continuous methods often compute this transformation through optimization of a cost function using standard optimization methods such as gradient descent. However, in the last years discrete optimization methods have become very popular. They turned out to be powerful optimization tools in a wide range of vision problems. The key contribution of this work is a general framework which bridges the gap between the continuous models used in deformable registration and discrete optimization algorithms. Additionally, we developed an optimization method based on the famous Alpha-expansion which overcomes the compromise in standard discrete labeling problems between computational speed and accuracy. We implemented a multi-level approach which is fast and capable to compensate for large and very small deformations at the same time. Furthermore, we introduce uncertainty estimation for non-rigid image registration. We are able to provide local uncertainty information for the recovered transformation. This allows us to evaluate the registration results but also offers new ways of advanced visualization. We tested our algorithms on optical flow estimation and medical image registration.     «

Non-rigid image registration is one of the most popular problems in computer vision. Many different approaches have been used to recover the transformation between two images. Continuous methods often compute this transformation through optimization of a cost function using standard optimization methods such as gradient descent. However, in the last years discrete optimization methods have become very popular. They turned out to be powerful optimization tools in a wide range of vision problems....     »