Predicting_Image_Rotations.md

Predicting Image Rotations

• Title: Self-training with Noisy Student improves ImageNet classification
• Publication: CVPR, 2020
• Link: [paper] [code]

Introduction

• Supervision learning is effective for feature learning, but there must be a lot of labeled data for successful learning.
  • Too expensive in reality, and requires an unrealistic scale
• Using unsupervised semantic feature learning is critical in processing massive data because it doesn't need any effort in labelling

What is "understanding image?"

• Requiring to understand the concept of the objects depicted in the image
  • Location in the image, their type, and their pose.
• In this thesis, Use rotation transformations to predict image.
  • It should be recongnized even in a rotated state.

Methodolgy

• F: ConvNet model, G: k distinct geometric variations (not random because they are discrete)
• g: Create converted image X (transformation applied)

• X | (theta): Transformation by changing the angle by theta for input X.

• Minimizing loss is our goal, so we can improve the direction in which loss decreases by comparing the calculations by rotating the angle!

Choosing gemometric transformation

• In thesis, model are trained by rotating Pi/2 rad.
• To successfully predict the rotation of an image, the model must learn to localize salient objects in the image.
• Recognize their orientation and object type, and then relate the object orientation with each type of object.

Comparing attention between supervised learning and self-supervised learning

• Self-supervised learning shows better attention maps than supervised learning.

Benefits of rotation

• computational cost similar to supervised learning.
• Learning speed similar to supervised learning.

Experiments

• Tests are based on RotNet(Network in Network).

- Multiple rotation blocks, performance best in the 1st or 2nd block above when comparing performance.

- Using four rotations is the best for accuracy.

- supervised NIN and RotNet + conv are the same architectures, and there is no significant difference in performance between supervised and self-supevised learning.

Reference

@article{DBLP:journals/corr/abs-1803-07728,
  author       = {Spyros Gidaris and
                  Praveer Singh and
                  Nikos Komodakis},
  title        = {Unsupervised Representation Learning by Predicting Image Rotations},
  journal      = {CoRR},
  volume       = {abs/1803.07728},
  year         = {2018},
  url          = {http://arxiv.org/abs/1803.07728},
  eprinttype    = {arXiv},
  eprint       = {1803.07728},
  timestamp    = {Mon, 13 Aug 2018 16:46:04 +0200},
  biburl       = {https://dblp.org/rec/journals/corr/abs-1803-07728.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}