Image Matting aims to separate a foreground object from an arbitrary image, which is explained as a combination of foreground and background. The result of a matting procedure is an alpha matte, which defines the influence of foreground and background on every pixel. The variety of matting approaches creates results of different quality. To measure the visual quality of an alpha matte, one can use subjective and objective evaluation methods. The human viewer plays a decisive role, as the human eye uses criteria for evaluation which may be even more important than the pixel-wise difference to the reference image. Subjective evaluations in the course of user studies are very time-consuming, thus automated calculations are necessary. These calculations will succeed, if they correlate with human perception. This motivates the use of methods that take the human visual system into account. In this work, a user study and automated calculations are conducted to infer the visual quality of matting results in the presence of three different classes of artefacts. These classes are spatial connectivity, gradient and possibly arising artefacts in foreground and background. Afterwards the results of study subjects and automated calculations are compared. The evaluation of the study has shown that objects with high connectivity and low gradient are classified as of high visual quality by experimentees. Traditionally used pixel-wise error measures do not correlate well with the visual quality as perceived by the study subjects. The assumption that artefacts in the background have a worse effect than those in the foreground, could not be confirmed by study subjects of the current study.
P. Rott: "Evaluierung von Fehlermetriken für Image Matting"; Supervisor: M. Gelautz, C. Rhemann; Institute for Software Technology and Interactive Systems, 2008.
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.