Archisman Bandyopadhyay

Subgraph-Based Methods for Rotation Synchronization

Rotation synchronization is a problem that occurs in 3D reconstruction problems. Given a set of relative rotations between pairs of images, which are modeled in a graph with the nodes as images and edges as 3D rotation matrices that may be noisy and corrupted, the problem is to find the correct absolute rotations for each image upto a global rotation. Our method builds upon cycle-edge message passing (CEMP), which gives estimates of the corruption or error levels of each relative rotation and then uses a minimum spanning tree (MST) with the error levels as the edge weights to get the absolute rotations. In our methods, we remove some edges of the graph based on certain criteria involving the error level of an edge relative to that of its neighbors, and then use CEMP+MST. We use two separate criteria in our work. We look at the accuracy of both methods compared to that of CEMP+MST on the whole graph for real image data. While our methods do not seem to consistently produce more accurate solutions, some ideas are proposed to pick better subgraphs for more accurate solutions.