19CVPR| Self-Supervised Representation Learning by Rotation Feature Decoupling

[XFeiF]

paper && code

The proposed method learns a split representation that contains both rotation related and unrelated parts.

They train neural networks by jointly predicting image rotations and discriminating individual instances.

In particular, the model decouples the rotation discrimination from instance discrimination, which allows to improve the rotation prediction by mitigating the influence of rotation label noise, as well as discriminate instances without regard to image rotations.

[XFeiF]

This work finds the disadvantages of RotNet, and proposes a feature decoupling method to handle it.

The feature decoupling can be divided into three parts as shown in the figure above.

1. Rotation classification. This part differs from RotNet as only part of the extracted feature f are used to recognize the rotation.
2. Rotation irrelevance. They enforce similarity between features of the same image with different rotation images. Formally, they minimize the distance between each feature of rotated images and their mean feature vector.
3. Image instance classification. Since features of rotated copies of an image have already been constrained to be close to their mean feature vector, they proceed to distinguish and spread out these mean feature. (This part may be hard to understand at least for me.)