athenoc's Journal


The project uses a bunch of techniques to rank all possible orientations of a image. The highest ranked orientation is predicted as the most suitable.


Some of the techniques used include : A Haar classifier (that can dynamically be built using already tagged images or a pre trained classifier is already available out of the box), a reinforced network with a reward based system that works by using user input (example, if a user manually changes an orientation after the network displayed it's predicted orientation), PCA analysis to understand the orientation of objects in the image, heuristics based on hues and colors in the image (Blues are likely to be on the top, solid objects are likely to be at the bottom) amongst others.


Check out the program in action down below!


Auto-Orient Images