Refinement Network for unsupervised on the scene Foreground Segmentation

In this paper we present a network for foreground segmentation based on background subtraction which does not require specific scene training. The network is built as a refinement step on top of classic state of the art background subtraction systems. In this way, the system combines the possibility to define application oriented specifications as background subtraction systems do, and the highly accurate object segmentation abilities of deep learning systems. The refinement system is based on a semantic segmentation network. The network is trained on a common database and is not fine-tuned for the specific scenes, unlike existing solutions for foreground segmentation based on CNNs. Experiments on available databases show top results among unsupervised methods.