Content-Aware-Rotation

• Reproduction of Kaiming He's Paper Content-Aware Rotation on ICCV 2013.
  
• I just test this project on Windows 10, there may be something wrong on Linux system.
  

Contents

1. Dependency
2. Formula Derivation
3. Usage
4. Results
5. References

Dependency

 Python 3.6 or newer

 pillow == 5.1.0

 numpy == 1.14.5

 opencv == 4.2.0

 matplotlib == 2.2.2

 tensorflow == 1.10.0

Formula Derivation

• Rotation Manipulation
  
  suppose that :
  ,
  then we have :
  
  further, we can get the derivative :
  
  

• Line Preservation
  
  suppose that :
  ,
  then we have :
  
  further, we can get the derivative :
  
  

• Shape Preservation
  
  suppose that :
  ,
  then we have :
  
  further, we can get the derivative :
  
  

• Boundary Preservation
  
  further, we can get the derivative :
  
  

• Optimization

• Step 1 : Fix θ solve for V
  Sparse Linear System :
  
  V is the minimum point of total energy above, thus, let :
  
  we can get the solution for V.
  

• Step 2 : Fix V solve for θ
  Part 1 : Fix Φ, update θ
  
  suppose that :
  
  we have :
  
  then let :
  
  by solving a sparse linear system, we can get the solution of θ :
  
  Part 2 : Fix θ, update Φ
  
  the solution of Φ can be easily approximately reached by enumeration based on iteration methods(increase β gradually).
  

Usage

• How to Use : download the whole project and run main.py
• folder ep : images which are used for formula derivation and some results.
• folder image : images which are used in original paper.
• folder lsd : line segment detection algorithm in python version.
• warp_mesh.py : image warping method realized by embedded function in tensorflow.

Results

• The Leaning Tower of Pisa 5.5°
  
  
  

• Palace Tower -6.1°
  
  
  

• House Building -5.8°
  
  
  

• The Oriental Pearl Tower 1.8°
  
  
  

References

Author : Kaiming He, Huiwen Chang, Jian Sun

•  Content-Aware Rotation----ICCV 2013
  

Matlab Version

•  iRotate