Salient Detection is the vision task of highlighting an object of interest. For me, this was the first time I’ve interacted with the use-case and it is effectively creating a silhouette of the object of interest. The more accurate the silhouette, the better the salient detector. It is an important but challenging task.
The paper notes there are broadly two branches of approach. Either bottom up or top down.
Bottom Up uses local feature contrasts to segment the image. To learn these local feature contrasts, both local and global features are extracted from pixels. As a result, they tend to leave un-contrasted maps.
Top Down is task oriented. It is using the prior knowledge of objects to aid in the segmentation of those objects, by essentially pruning out the bottom up feature contrasts that are predicted to not be a part of the analyzed object.