Computer Vision News - April 2018

This work is about a new loss function that can work with non-aligned data . Its applications are several with the belief that it is applicable to more tasks than are showed in the paper. The main novelty being presented here is, according to the authors, the first loss function that can handle non-aligned data in a training process of a feed forward CNN ; that is, it can train the net with source and target images which are nonaligned. Classic loss function such as L1 or L2 cannot handle it because they assume pixel-to-pixel correspondence. What the authors propose is a new loss function that can handle this non-alignment. In the authors’ previous paper presented at CVPR 2017 , they researched the task of template matching, trying to find a good similarity measure between images. The proposed measure named DDIS achieved state-of-the-art results; however, it does not have a meaningful derivation and thus cannot be used as a loss function. To find How to make it derivable, so that it can be used as a loss function is the goal of the discussed paper. The key idea of the algorithm is to represent the images as a set of points in a high dimension space. Then in order to solve the challenge the distances between all points are used. A statistical measure is then taking place over this affinity matrix between the two sets of points representing the images. What can be built on these findings: i) if people tried to use this loss function for other tasks; ii) others can design even better loss functions that follow this statistical direction which has some very interesting properties. This loss function has nice applications: for example, a single image is animated using a target video; it's used for style-transfer which is showed in many papers, but according to the authors they have a new kind of style-transfer which transfers the style semantically. For example, if we take a content image of Trump Research 30 Research Computer Vision News The Contextual Loss for Image Transformation with Non-Aligned Data