Rotation Representations in Deep Learning
In this project, we reviewed several popular rotation representations, discussed their discontinuitiesand other problems in a direct regression task. Different from previous works, we focused on giving an intuitionon why those representations are discontinuous and their respective properties in a regression task. Totest our hypothesis empirically, we generated a dataset based on ShapeNetV1 3D model dataset anddesigned experiments on 3D rotation estimation task using CNN as model and images as input. We triedseveral traditional rotation representations as well as the 6DOF rotation representation proposed in and we found that 3D and 4D rotation representations indeed yield larger training and testing errorsdue to their discontinuity, while rotation matrix works better and the 6DOF representation works bestin the 3D rotation pose estimation task.
[This is my final project for 16-720: Computer Vision at CMU]