Off-line Foveated Compression and Scene Perception: An Eye-Tracking Approach

With the continued growth of digital services offering storage and communication of pictorial information, the need to efficiently represent this information has become increasingly important, both from an information theoretic and a perceptual point of view.

There has been a recent interest to design systems for efficient representation and compression of image and video data that take the features of the human visual system into account. One part of this thesis investigates whether knowledge about viewers' gaze positions as measured by an eye-tracker can be used to improve compression efficiency of digital video; regions not directly looked at by a number of previewers are lowpass filtered. This type of video manipulation is called off-line foveation. The amount of compression due to off-line foveation is assessed along with how it affects new viewers' gazing behavior as well as subjective quality. We found additional bitrate savings up to 50% (average 20%) due to off-line foveation prior to compression, without decreasing the subjective quality.

In off-line foveation, it would be of great benefit to algorithmically predict where viewers look without having to perform eye-tracking measurements. In the first part of this thesis, new experimental paradigms combined with eye-tracking are used to understand the mechanisms behind gaze control during scene perception, thus investigating the prerequisites for such algorithms. Eye-movements are recorded from observers viewing contrast manipulated images depicting natural scenes under a neutral task. We report that image semantics, rather than the physical image content itself, largely dictates where people choose to look. Together with recent work on gaze prediction in video, the results in this thesis give only moderate support for successful applicability of algorithmic gaze prediction for off-line foveated video compression.