Document worth reading: “Deep Learning for 2D and 3D Rotatable Data: An Overview of Methods”

One of the reasons for the success of convolutional networks is their equivariance/invariance beneath translations. However, rotatable data equivalent to molecules, residing cells, frequently objects, or galaxies require processing with equivariance/invariance beneath rotations in situations the place the rotation of the coordinate system would not affect the which means of the information (e.g. object classification). On the other hand, estimation/processing of rotations is crucial in situations the place rotations are important (e.g. motion estimation). There has been newest progress in methods and thought in all these regards. Here we provide an overview of current methods, every for 2D and 3D rotations (and translations), and decide commonalities and hyperlinks between them, throughout the hope that our insights is likely to be useful for deciding on and perfecting the methods. Deep Learning for 2D and 3D Rotatable Data: An Overview of Methods