We have presented the importance filtering algorithm for content-aware image retargeting. It directly uses the original image as the constraint to filter and estimate pixel importance so that it is consistent with the image structure. This is the key to minimize the visual distortion and yet preserve the prominent image contents. The constraint is applied on the gradient of pixel shift, instead of directly on pixel shift. This further avoids undesired distortion such as pixel swap that occurs to many earlier methods. The importance filtering operations are highly efficient and ready for real-time applications. We also show that easy extension to video re- targeting is promising.
One potential improvement to the importance filtering algorithm is to extend the one-dimensional shift gradients to 2D. Even though the pixels all shift along the same dimension, the shift-map on the 2D image has a 2D gradient field. We are developing methods to estimate such 2D shift gradients and then optimize their integration to construct the shift-map by methods such as alternative 1D filtering or Poisson blending. We believe this will further improve the 2D smoothness and consistency of the resized image.