Computer Vision News - May 2019

Image Segmentation: Image segmentation into foreground and background regions can be conceived of as decomposition of the image into a foreground layer y1 and background layer y2, combined according to a binary mask m, which for every pixel x follows the formula: This formulation naturally suits the framework proposed by the article, which defines a ‘good image segment’ as one that is easily put together using parts that belong to it, but hard to put together using parts from other segments of the image. To encourage the Segmentation Mask to be binary, we use the following loss term: Double-DIP can achieve high quality segmentation based solely on layer decomposition, with no supervision, as illustrated in the figure below: Many more results can be viewed on the project’s website. It’s true that there are many other approaches to segmentation (among them, semantic segmentation), that perform even better than DIP. However, they all suffer from the disadvantage that they need to be trained on large datasets. 9 Research Computer Vision News ( ) = ( ( ) − 0.5 ) −1 ) ( ) = ( ) 1 ( ) + (1 − ( )) 2 ( Double-DIP