Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Metrics for quality assessement of watermarked 3D polygonal models

: Prasiswa, J.
: Funk, W.

Darmstadt, 2004, 70 pp.
Darmstadt, TU, Dipl.-Arb., 2004
Fraunhofer IGD ()
3D polygonal models; perception; watermarking

Non-perceivability is of great importance for digital watermarking, thus it would be desirable to obtain objective measures of visual quality. These measures are needed in simplification of 3D models and in compression as well. In an image-based and an animationbased experiment, human perception of distortions caused by the Sculpting Based Watermarking algorithm by Benedens was recorded. Two experimental methods were utilised: 'two alternative choice testing' and a rating test based on the 'double-stimulus impairment scale method'. An online evaluation method for the experiments was developed, it proved to be effective and convenient for both the participants and the experimenter. For the configurations chosen in accordance to the default configuration the watermarking algorithm proved to be non-transparent for both experimental designs. The perception rates were higher for still images than for the animated sequences. Results also revealed a high dependency of the perceived quality on the chosen model and the key used during watermarking. Furthermore, several metrics were introduced and their ability to predict perceived quality was evaluated. The metrics differed in effectiveness depending on the choice of stimuli. The image based Lindstrom metric was effective for still images, much less for animations. The best overall results were achieved by two geometric metrics, the symmetric Hausdorff distance giving the maximal distance between original and marked mesh and the root mean square Hausdorff distance. These results confirm previous experiments, where geometric metrics also proved to be effective for small quality differences. An accurate correspondence between perceived and objective quality could not be established as results varied between models.