Computing That Serves

Advancements in Object Selection in Images


Thursday, March 9, 2017 - 11:00am


Brian Price


Bill Barrett

Selecting objects in images is an essential step in many image editing and special effects workflows.  This talk discusses advancements being made in image object selection.  It will touch on work that has made it into Adobe products that are used by professionals across the world, but will mainly focus on research efforts to greatly simplify object selection.  A key limiting factor in interactive object selection is that most methods rely largely on color and edge information when computing a selection.  However, this information is often ambiguous due to overlapping foreground and background color distributions, weak edges, and complex textures and structures in the object and background.  To address this, we use a dataset of images with previously-generated object masks to provide more knowledge about the objects and backgrounds.  We first explored retrieval-based methods where for a given image a number of images that are similar in appearance are retrieved and used to transfer the selection to the target image.  Given the rise of deep learning methods, we then transitioned to using convolutional neural network approaches to directly compute the selections, giving us state-of-the-art results.  Both of these approaches will be discussed as well as their usage for automatic segmentation as well as interactive segmentation given different input schemes. 


Brian Price is a Senior Research Scientist 2 in the Imagination Lab in Adobe Research. He actively researches computer vision problems such as segmentation and matting, document analysis, color harmonization, image hole filling, and image generation to name a few. He developed technologies that have gone into many Adobe products including the Select and Mask tool in Photoshop, the Smart Selection and Auto Select tools in Photoshop Mix, and the Refine Edge and Key Cleaner features in AfterEffects. Prior to coming to Adobe, He received his PhD degree in computer science from Brigham Young University.