People     Research     Contact CS Department Yale

Understanding and Improving the Realism of Image Composites Su Xue Aseem Agarwala Julie Dorsey Holly Rushmeier Yale University Adobe Systems, Inc. Yale University Yale University Abstract Compositing is one of the most commonly performed operations in computer graphics. A realistic composite requires adjusting the appearance of the foreground and background so that they appear compatible; unfortunately, this task is challenging and poorly understood. We use statistical and visual perception experiments to study the realism of image composites. First, we evaluate a number of standard 2D image statistical measures, and identify those that are most significant in determining the realism of a composite. Then, we perform a human subjects experiment to determine how the changes in these key statistics influence human judgments of composite realism. Finally, we describe a data-driven algorithm that automatically adjusts these statistical measures in a foreground to make it more compatible with its background in a composite. We show a number of compositing results, and evaluate the performance of both our algorithm and previous work with a human subjects study. Paper To appear in Siggraph 2012 (ACM Trans. on Graphics) [PDF] Supplemental Materials A. Likelihoods of offsets of statistical measures (Section 2) All likelihoods we examined are shown, a subset of which are shown in Figure 2 in Section 2. B. Stimuli and results of MTurk experiment (Section 3) We show all stimuli images, the realism ratings, and the fitted Gaussians described in Section 3. C. Offsets of different zones in natural images (Section 4) Figure 8 in Section 4 only shows the distributions of offsets for luminance. Here we show the distributions for CCT as well as saturation. D. Complete compositing results (Section 5) We show all the results adjusted by five methods over 48 images, which are the stimuli for user evaluation in Section 5. Data The filtered subset of LabelMe. (Section 2, 4) 4126 images; with meaningful, un-occluded foreground objects in front of background scenes. In Section 2: identify statistical measures. In Section 4: train the classifiers. Stimuli of image on human realism ratings. (Section 3) · 20 images with precise mattes · Three matrices for each image: luminance, CCT, saturation. (7x7 manipulated images in a matrix) Last Updated  2012-02-22 15:45:12