Predicting Rotation

Learn to implement self-supervised learning via rotation prediction.

Rotation prediction

As illustrated in the figure below, the critical idea of rotation-based self-supervision is to rotate an input image and then ask the neural network to predict which rotation degree was applied to the picture. Generally, this pretext task is posed as a 44-way classification where the input image, XiDsourceX_i \in D_{source}, is randomly rotated by r{0°,90°,180°,270°}r \in \{0\degree, 90\degree, 180\degree, 270\degree \} degrees and then passed through the neural network, f(.)f(.) for prediction. The rotation prediction loss is defined as:

Get hands-on with 1200+ tech skills courses.