Skip to content

done1892/Advanced-Machine-Learning-Project

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

73 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

CAOS - Continuous Angle Orientation System

The task of image orientation consists in the prediction of the most natural angle that a human would use to take a picture. In this paper we propose CAOS, a new system to allow a common convolutional neural network to deal with the cyclical nature of predicting the natural angle of an image. The solution is not straightforward due to the periodic behavior of angles. The model proposed is a supervised neural network that doesn't need feature points or landmarks as inputs but uses a standard convolutional neural network with a particular kind of loss function and output pair. The model reaches state of the art results in a specific domain without signs of overfitting.

  • The PDF file contains the report about this work.
  • The ipynb file contains Pytorch implementation of CAOS, described below.

    The first part of the notebook includes initialization of functions which will be used in dataloader to perform geometric transformations, such as rotations, cropping, etc. After, a Pytorch dataloader has been instantiated, it can handle one channel images and performs preprocessing and data augumentation, and after that feeds the neural network. The pretrained NN used for this task is SqueezeNet which outputs an array of two elements. The network has been trained through stochastic gradient descending using the Adam Optimizer. The Loss function used is customized and computes the chord between the angle predicted by the network and its real value. Gradients optimize this function in the training loop for 80 epochs. In the last part the trained model has been used for inference, and evaluation on some images

    For further details please open PDF file, which contains the report on the work.

    • You can see an example of CAOS in action below

  • A bunch of inference results on famous AI scientists.

Authors:

About

CAOS - Continuous Angle Orientation System

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 3

  •  
  •  
  •