CVPR Daily - 2019 Thursday

Many apparently unrelated computer vision tasks can be thought of and dealt with as special cases of decomposition into separate layers. To name just a couple of prominent examples: • Image segmentation -- which can be defined as decomposition into areas belonging to a background layer and areas belonging to a foreground layer. • Image dehazing -- which can be defined as decomposition into a clear, clean image and the dehazing map layer. The authors propose a unified framework for unsupervised layer decomposition of a single image, based on Deep-image-Prior (DIP) networks. Deep-image-Prior (DIP) networks, introduced at CVPR 2018, are a type of generative network that learns the low level statistics of a single image -- is trained on a single image. In the article, the authors show how stringing together several DIP networks provides a powerful tool for decomposing images into their basic elements -- for a wide variety of tasks. The authors believe this versatile applicability derives from the fact the internal statistics of a mixture of layers is more robust and has better representation capabilities than each layer separately. The authors show the applicability of their approach to a variety of computer vision tasks, including watermark-removal, Fg / Bg segmentation, image dehazing and transparency decomposition in video images, among others. All of these capabilities are achieved when the network is trained on a single image with no additional data provided. A unified framework for image decomposition -- below are illustrations of the article’s approach in action. Three different tasks redefined as decomposition of the original image viewed as a mixture of simpler basic layers. This approach of image decomposition into a number of basic layers -- provides a unified framework for dealing with a wide number of apparently disparate and unrelated computer vision tasks. by Assaf Spanier This review was first published on Computer Vision News of May 2019: Unsupervised Image Decomposition via Coupled Deep- Image-Priors. We are indebted to the authors (Yossi Gandelsman, Assaf Shocher and Michal Irani), for allowing us to use their images. The paper is here. If you would like to find out more about this work, come along to their oral [3.2C] today at 13:53 and poster [151] at 15:20-18:00. 6 DAILY CVPR Thursday Presentation