"Image Matting" is the process of separating a foreground object from a background in digital images. In order to quantitatively compare the results of matting algorithms, a benchmark test, which is based on a dataset of reference solutions (ground truth), is required. Such a benchmark should rely ideally on a large number of test images. However, unoptimized research implementations of novel matting algorithms oftentimes are not capable of processing such large datasets in reasonable time. Thus, a major challenge when designing a web-based benchmark test is the selection of a suitable image set of manageable size.
The goal of this diploma thesis is to select a representative subset from a recently introduced ground truth dataset of 61 images that can then be used for the comparison of matting algorithms. The challenge is to select a small subset of images that maintain, as far as possible, the properties of the full set. This was achieved as follows. First the im-ages were categorized according to the boundary characteristics of the foreground objects. Afterwards, for each image, alpha mattes were computed using various matting algorithms. The error rates of those mattes were determined by the pixel-wise comparison of the resulting alpha matte with the reference solution. Finally, the images with the high-est error rates from every category were selected for the compact dataset, since they represent the largest degree of difficulty for the matting algorithms. The calculations experimentally confirm that the selected subset has indeed similar properties as the full set.
A further aim of the diploma thesis was to prepare the results of this work for didactic usage. For that purpose, an online course for an E-Learning platform was created with modern communicative devices. The online course is organized into modules that allow the self-organization of learning processes. In the first major module, the participants are introduced into image matting and important matting methods are presented. The second major module aims to interactively present the results of this work. For instance, a user interface was implemented that allows the users to rank different matting results according to their visual quality. In the future, the results of such a ranking could be used to evaluate to which extent error measures for image matting correlate to the visual quality.
J. Pucher: "Erstellung eines Standarddatensatzes für die Evaluierung von Alpha Matting Algorithmen und Aufbereitung der Resultate für die Lehre"; Supervisor: M. Gelautz, C. Rhemann; 188-2, 2010.
Click into the text area and press Ctrl+A/Ctrl+C or ⌘+A/⌘+C to copy the BibTeX into your clipboard… or download the BibTeX.