Implementation of a Graph-Convolutional Net for Rotation Synchronization
Rotation synchronization, also known as rotation averaging, is a fundamental problem in computer vision that aims to recover absolute orientations from relative rotations between pairs of cameras. It has many applications such as determining camera orientation using global Structure-From-Motion (SfM) algorithms and pose graph optimization in visual SLAM or sensor networks. Current approaches to rotation synchronization can be divided into two categories: optimization-based and deep learning-based, with optimization-based being far more popular. The goal of my project was to take a deep learning approach toward rotation synchronization and implement a graph-based convolutional neural network that produces accurate absolute rotations despite corrupt information. The results of this study suggest the efficacy of deep neural networks in the rotation synchronization problem is promising.