We describe the multi-reference alignment (MRA) method to solve a class of problems, including registration of noisy signals over the d-dimensional sphere and orientation estimation of noisy cryo-EM images. The main feature of the MRA method is that we simultaneously optimize over all data points with respect to the noise model. We provide the MRA solutions to the registration problem in 1-dimension and 2-dimension, and the cryo-EM problem. In the process, we will have to discretize SO(2) and efficiently discretize SO(3). We will demonstrate, through numerical simulations, MRA outperforms classical methods such as cross-correlation, SO(2)-synchronization, and invariants for the registration problem in 1-dimension. We will possibly include comparisons against SO(3)-synchronization in 2-dimension as well.
Estimating group transformations via semidefinite programming
Yutong Chen - PACM Graduate Student
Apr 15 2015 - 3:00pm
110 Fine Hall