Antialiasing for Automultiscopic 3D Displays

Matthias Zwicker, University of California, San Diego
Wojciech Matusik, Mitsubishi Electric Research Laboratories, Cambridge
Fredo Durand, Massachusetts Institute of Technology, Cambridge
Hanspeter Pfister, Mitsubishi Electric Research Laboratories, Cambridge

This is supplemental material for the EGSR 2006 paper "Antialiasing for Automultiscopic 3D Displays", including a Java applet and four animated video clips.

Java Applet

We provide a Java applet to explore optimal sampling and filtering for 3D image acquisition and display. The applet requires the Java Runtime Environment (JRE) 1.4 or newer. The applet can be found here.

Description

The user specifies a desired scene that he would like to acquire. The applet computes the minimum sampling rate such that replicas of the input spectrum exactly touch the display prefilter, but do not overlap with it. This guarantees optimal display quality. The derived sampling rate consists of the number and spacing of cameras, and their resolution and field of view. This corresponds to the scenario described in Section 6 in the paper.

Usage

Animated Clips

The four animations below compare full resampling with a combined reconstruction and display prefilter to resampling with only a light field reconstruction filter. The animations show simulated views of a 3D display with 8 views (i.e., 8 view-dependent subpixels) under a horizontal left-to-right translation of the viewpoint.

Aliased Views

Buddha, Elephant

These animations use only a light field reconstruction filter, but do not include the display prefilter. This leads to strong aliasing artifacts in the directional domain, i.e., along the v-axis in the two plane parameterization of the display. Aliasing becomes apparent as temporal patterns in the animations.

Properly Resampled Views

Buddha, Elephant

These animations use the combined resampling filter, which includes the display prefilter. The combined resampling filter largely avoids aliasing artifacts. However, some temporal patterns are still visible because of the use of Gaussians, which are not ideal low-pass filters. Observe the shallow depth of field of the views imposed by the display bandwidth.