Down the VR rabbit hole: Fixing judder
blogs.valvesoftware.com
blogs.valvesoftware.com
"I was initially skeptical of the importance of low persistence, preferring a push for 120hz, but it is a BIG DEAL."
I'm not so sure. On the first post in the sequence, someone had left a comment about DLP, which has extremely fast switching times. This was agreed to be impractical on the grounds that DMD pixels are either on or off and can only achieve intermediate levels via PWM (which takes enough time to do).
I initially didn't think much about it, but it festered in my head for a few hours, and I started researching it a bit. I think it could actually work.
The thing is, our eyes are expecting continuous input, yes, but they don't actually have the bandwidth to really pull in anywhere near a thousand 24 bit images every second. The problem is the quantization, not the actual amount of information. Obviously a DMD is not going to fix the spacial quantization, but I believe it can fix the quantization in time.
The DMD I was looking at can be switched by its controller at 2.9 kHz for arbitrary input, and 4.2 kHz for pre-loaded input. From what I can tell, there's other overhead going on there, and it looks like you should be able to get it up to about 5.5kHz if you built a custom controller.
What I propose is to run the display at 5500 fps... at 1 bit per channel. Hear me out!
If we set the threshold for detecting small variations in light at 200Hz (this is what is recommended for PWM dimming of LEDs, for example), then that gives us about 28 pulses to work with, within the eye's tolerance. Now we have some options. Maybe the cost of DMDs goes way down in the next few years. In that case, we could stack them (increasing the new LEDs' intensities by powers of 2) and get exponentially more levels with each DMD. As it stands now, that would get expensive very quickly - with only one for each eye, the QVGA version of this is already at about $200.
Of course we do want colors, so we'd divide up our time between the RGB channels. Give, say, 6 levels to blue, 8 to red, and 14 to green.
So by using temporally dithered 1-bit frames, we should be able to get (as far as our eyes can tell) continuous motion (but still quantized to pixel boundaries). Now, thinking at the 5ms scale instead of the 200μs scale, we should be able to do more dithering. You'd be surprised at how much you can get away with even with purely spatial dithering with only 6.5 bits per pixel if you divide those bits up right.
Edit: I realized I got mixed up between my modes of thinking. The controller doesn't drive the DMD full speed because it leaves gaps between each channel. When I worked out that 5.5 kHz figure, I was still thinking of at least one DMD per channel. So you get even fewer colors than I thought.
One thing I forgot to mention though, is that as far as I can tell from these data sheets, the bottleneck by far is bandwidth. Having studied it for a while, I'm not entirely sure that we don't already have the technology to build a single-DMD full color 1kHz DLP. There is however no doubt that you could build a full color 1kHz DLP using multiple DMDs. I'm just not sure it would fit in an HMD.
Such a display would be made of pure cash, of course. Can't have everything.
Aside, this stuff is really fascinating, and it's great to see the field of human-computer interaction pushing up against formerly unknown biological phenomena.
http://www.youtube.com/watch?v=ZuSUHuSceYc
The tech's called Lumipen, and they use something called Saccade Mirrors to do it.
It would not fix smooth persuit tracking of a moving object while the head is stationary though, as you could not selectively shift the moving object to match eye rotation while also keeping the background stationary.
[1] http://www.physikinstrumente.com/en/products/prdetail.php?so...
Animation demo of eye-tracking based motion blur: http://www.testufo.com/#test=eyetracking
Animation demo of strobing to reduce motion blur: http://www.testufo.com/#test=blackframes
View these links in Chrome or another web browser that supports perfect VSYNC animations. The pages will attempt to detect if your browser supports VSYNC. IE10+ works well (up to 100Hz), Chrome works well (up to >144Hz), and Safari6+ (iPad's work too). FireFox 22 will judder too much; FireFox 24+ pre-beta adds VSYNC support -- I am the person who convinced Mozilla to add native support for 120fps animations as Bugzilla@Mozilla #856427; now in FF24 pre-beta.
Mark Rejhon (Blur Busters Blog)
If you're running "ToastyX Strobelight" (easy lightboost programming utility) on a 120Hz LightBoost-compatible monitor -- hit Control+Alt+Plus to enable LightBoost strobing, then Control+Alt+Minus to disable LightBoost strobing. Control+Alt+1 will program the strobing to 1.4 milliseconds, while Control+Alt+0 will program the strobing to 2.4 milliseconds (oscilloscope measurements). The 1.4ms-vs-2.4ms is actually noticeable in the castle at the top of http://www.testufo.com/#test=photo&pps=1440
It's a very interesting exercise in hacking LightBoost 2D as a programmable-persistence computer monitor (100% software hack; no hardware modifications needed; no 3D kit needed)