Speaker: Johannes Kopf , Microsoft Research

Tuesday, April 15th, 13:00, in BC 329

Abstract:
"Guided Image Completion" --- recent work on improving image completion by combining a low-level patch-based algorithm with mid-level structural cues, as well as predicting result quality before the completion is carried out. "Image-based Rendering in the Gradient Domain" --- we show how IBR for complex scenes can be improved by rendering the novel image in the gradient domain followed by a image domain Poisson reconstruction. Our algorithm is able to handle general scenes including reflections and similar effects without explicitly separating the scene into reflective and transmissive parts, as required by previous work. "Content-Adaptive Image Scaling" --- recent work on downscaling images by optimizing the shape and locations of downsampling kernels in a content-adaptive manner to better align with local image features. We optimize these kernels to represent the input image well, by finding an output image from which the input can be well reconstructed. This enables our algorithm to produce sharper results (without ringing artifacts) and preserve image features better than existing content-agnostic scaling filters.