How far does SSS have to move the sensor (warning - very long)

part 2 - test and results

Essentially I took twenty shots of a some lined white paper with the lines arranged vertically. Ten were taken with SSS enabled, ten without, The paper was mounted on a wall and illuminated fairly evenly. The photos were taken with the paper approximately 5 metres from the effective optical centre of the lens (that is about 5.5 metres from the focal plane). The exact distance doesn't matter much, but it serves to ensure that the lens was indeed behaving like a 500mm (many extreme zooms don't honour their nominal focal length at very short focussing distance). I chose an exposure time of 1/25th second at f8 using spot AF. I took the photo handheld without any special bracing or support.

I then processed the raw files into TIFF and used Photoshop and the pixel counting setting under “info” to assess the horizontal “smearing” of the vertical lines with and without SSS. A tripod image revealed that the line width was (to the nearest) 5 pixels. Note that assessing the “smearing” needs a degree of judgement as the border is not completely sharp and that one side or other tends to fade out to grey according to which way the camera moved. The results I found were the following

With SSS enabled

Average “error” was 2.6 pixels with a standard deviation of 4.1 pixels. One “outlier” result of 12 pixels pushed the average up from about 1.4 pixels so mostly the performance was a lot better.

Without SSS

Average “error” was 30.9 pixels with a standard deviation of 19.1. The “minimum” deviation was 3 pixels with a maximum of 61 pixels.

As the A700 has a sensor pitch of 5.4 microns, and we take the normally accepted value of the limit of the Circle of Confusion for A700-sized sensors as 20 microns (3-4 pixels) then most of the stabilised shots fall within that limit (two were outside) whilst all but one of the unstabilised shots exceeded it. Of course I only measured horizontal deflection, so there will also have been vertical deviations too so this over-emphasises the effectiveness of the SSS. Also the pixel-peepers seem to expect images to be sharp at the pixel level (which is beyond the normal DoF definitions used historically).

Taking the maximum unstabilised deflection that I saw of 61 pixels, this is equivalent to a deflection of 0.33mm. This deflection of 0.33mm at 500mm is equal to an angle of just over two minutes of arc (which is about the accepted resolution limit of the eye). The 50mm pistol event at the Olympics has a bull which subtend an angle of about 3 minutes of arc (note, I’m not claiming great accuracy – I didn’t have to aim, but the numbers serve to indicate that the results aren’t wholly ridiculous).

What does this mean? As far as the full-frame issue is concerned, as long as the image circle is at least 1mm larger than the sensor all round (deflection can occur in both X & Y axis so could be 1.4 x at the image corner) then this would suffice to compensate for the same deflection using a 1,000mm lens. Even if there is no margin to spare, we would only have to sacrifice about 2mm of the “long” side of the full frame to give this margin (or a crop factor of about 1:1.05).

I don’t really know how easy it is to accelerate a sensor to cope with movements of 0.33mm in 1/25th of a second, but one thing is very obvious. It has to be incredibly precise and fast. From what I can see it is typically better than 2 pixels or about 10 microns (1/100th of a millimetre). Computer hard drives do a great deal better, but they only move with one degree of freedom. They also don't have to include gyros/accelerometers and use intelligence to distinguish tremors from panning.

Overall, I’m pretty happy that sensor based stabilisation can cope with both long lenses and full-frame for “reasonable” requirements. Maybe beyond 1,0000mm something different will be required, but that's very specialised and there is nothing stopping Sony using lens-based stabilisation as well should they ever choose (just automatically disengage SSS).

For the record, I think that the future will eventually be very high definition EVFs (with localised magnification for detailed manual focussing). That will fix the Viewfinder stabilisation issues, although the purists will hate it. The mirror and all the other kludgy mechnisal stuff can then go and we can move to electronic shutters and moving images (if they can fix the autofocus issue).