Jaggies testing theory

The three Jaggies test evaluate how good a video processor is at deinterlacing objects that are moving. The first "jaggies pattern" is a simple spinning bar, the second "jaggies pattern" is a set of three bars with a waving motion (like waving your hand), and the third test is a real-world picture of the American flag in the wind. These all represent scenes that a deinterlacer will easily be able recognize as having motion.

ATI

NVIDIA

XGI

In the ATI interview, I hypothesized that they were using some form of edge-adaptive deinterlacing when talking about "vector-adaptive" anti-aliasing. ATI is reluctant to discuss this any further. On the other hand, XGI advertises that they use the edge adaptive deinterlacing. Remember that edge-adaptive interlacing works in this way:



Suppose that you have these pixels at the odd rows 1 and 3:





The video processor has to figure out what 5 pixels should go in between





A simple bob would just average the pixel above and below for each.





Resulting in this:





A better intrafield method of deinterlacing is edge-adaptive diagonal interpolation in which you examine several pixels from the row above and below the discarded pixel instead of just the pixels above and below. If an edge is detected, then the missing pixel is interpolated by averaging the pixels along that edge. So, first the video processor looks for edges in several directions





When it detects the edge, it interpolates along that axis.





You can then use the nearest pixel outside of the edge resulting in





The difference doesn't looks like much until you take a step back and compare the two. With traditional bob, you get artifacts that result in jaggies but with edge-adaptive directional interpolation it looks exceptionally smooth. (Try squinting for the maximum effect).





Amazing isn't it? Diagonal filtering was a feature pioneered by Faroudja. The next question is how do I know if the correct "edge" is the red part





or if the background happens to be red and the white line is what I'm supposed to interpolate?





Resulting in :





Was the picture supposed to represent a red slash on a white background or a bold white backslash on a red background? It's impossible to say. This is why deinterlacing isn't a brute-force computational solution. There's an art to video processing. You could have two video processors with diagonal interpolation, but one might be more accurate at guessing the appropriate direction. Likewise, you could have an exceptionally exotic deinterlacing strategy that only worked in some conditions.