Gaze-contingent visual communication
Andrew T. Duchowski · 1998
Virtual environments today lack realism. Real-time display of visually rich scenery is encumbered by the demand of rendering excessive amounts of information. This problem is especially severe in virtual reality. To minimize refresh latency, image quality is often sacrificed for speed. What's missing is knowledge of the participant's locus of visual attention, and an understanding of how an individual scans the visual field. Neurophysiological and psychophysical literature on the human visual system suggests the field of view is inspected minutatim through brief fixations over small regions of interest. Significant savings in scene processing can be realized if fine detail information is presented just in in a gaze-contingent manner, delivering only as much information as required by the viewer. An attentive model of vision is proposed where Volumes Of Interest (VOIs) represent fixations through time. The visual scanpath, composed of raw two-dimensional point of regard (POR) data, is analysed over a sequence of video frames in time. Fixation locations are predicted by a piecewise auto-regressive integrated moving average (PARIMA) time series model of eye movements. PARIMA model parameters are derived from established spatio-temporal characteristics of eye movements. POR data is fitted to the PARIMA model through the application of the three-dimensional wavelet transform. Identified fixations are assembled into volumes in three-dimensional space-time, delineating dynamic foveal attention. The attentive visual model utilizes VOIs to synthesize video sequences matching human visual acuity. Specifically, spatial resolution drops off smoothly with the degree of eccentricity from the viewer's point of gaze. Seamless degradation of individual video frames is accomplished through inhomogoneous wavelet reconstruction where the intersections of VOIs and frames constitute expected foveal regions. Peripheral degradation of video is evaluated through human subjective quality testing in a gaze-contingent environment. The proposed method of visual representation is applicable to systems with inherently intensive display requirements including teleoperator and virtual environments.