K-7 and its continuous autofocus -- PART I

Probably the most often criticised feature with a Pentax dSLR is the speed of its SAFOX VIII autofocus system. Interestingly, its accuracy is believed to be very good (which matters most for many situations). But two domains remained to be improved:

• Speed and availability of low light autofocus (AF-S).
• Speed and hit rate of continous autofocus (AF-C) for fast moving subjects (aka action or sports).

In this blog post, I'll study the latter: AF-C. This will be a two part article:

PART I: A car moving towards the photographer

I decided to follow the methodology of a test which was carried out about a year ago by German magazine fotoMAGAZIN (published in issue 5/2008). At the time, I posted a summary about this here:

Magazine Review: K20D/SDM AF.C performance vs. 8 competitors

The method is as follows:

• Continous shots (burst, continous AF aka AF.C) of an approaching white Chrysler LeBaron, tripod mounted and panned centering onto radiator grill. Photograph besides road. Road straight. Test done on a straight highway.
  Original: red Chrysler PT cruiser, closed military airport.
• Controlled speed: 100 km/h (62 mph). Starting photo burst at 100m distance (109 yards, 328 ft). Stopping when passing by. Time to drive through 100m: 3.6s.
• Focal length: 35mm equivalent of 200mm (DA 18-250 at 142.5mm/5.6).
  Original: DA SDM 50-135/2.8 at 135mm/2.8
• Aperture: f/5.6
  Original: f/2.8
• Shutter: 1/180 or shorter
  Original: 1/320 or shorter
• Light: Bavarian white blue sky. Changing conditions.
  Original: Daylight overcast with grayness. Constant conditions.
• Storage: JPEGs onto Sandisk Extreme III.
• Statistics: 6 series, averaged.

So, some conditions had to be changed. Not necessarily in favour of the new setup (assuming that the DA SDM 50-135/2.8 is superior to the DA 18-250).

At f/5.6, depth of field (DoF) is larger than at f/2.8. E.g. at 50m distance, it is 15.7m for f/2.8 and 33.9m for f/5.6. This may look like a big difference. However, the AP-C circle of confusion (20µm) is 4 pixels large and watching for 2px wide sharp edges means that we can compensate for this. Let's just require that 100% crops look pretty sharp.

To have a point of calibration, I decided to test both, a K20D and a K-7.

Disclaimer: I later discovered that I made the following mistakes in running the test. It doesn't ruin the results but should be taken into account: Fine sharpness set to 3 (too high); ISO in some runs set to 1600 (too high), others to 200 (too low); JPG noise reduction not turned off; JPG quality set to 4 stars (too high); driver not driving exactly constant speed.

The problem with 4 star JPG quality is that not 38 (K20D) or 40 (K-7) images are shot in hi-burst but only about 17. The K-7 can (theoretically) shoot 19 images within the time a car needs to travel 100m (3.6s). More if one launches too early or the car is too slow.

Here are my results:

K20D results:

5 runs, per run (total=++/+/o/-):

1. 13=8/3/2/0 too slow
2. 12=6/4/1/1 too far (150m), speed ok)
3. 13=7/0/6/0 too slow
4. 12=9/2/1/0 little too slow and little too far
5. 14=7/5/1/1 too slow and little too far

Fig.1: The 100% crop images for run #2 are all shown in the compound image above and these are the ratings: ++/+/++/++/++/+/++/++/+/+/o/- (last one was too near to have any chance, and last one is not a crop).

The magazine's original result was like

• 7.8=5.8/2 (meaning 5 or 6 good ones and always 2 bad ones; 25.6% bad; rank 9/9)

Taking run #2, discarding the first 4 images (farer than 100m) and counting 2/4/1/1 as 6/2 we would roughly get at the same result. So, I decide to count the first two qualities ++/+ as "good".

In an attempt to simplify my analysis, I then get my new result:

• 12.8=10.2/2.6 (20.3% bad; rank 8/9)

K-7 results:

Due to its higher speed, the K-7 fires away 14 to 20 shots. It slows down towards the end. Maybe because of AF.C needing more time, maybe because the buffer fills up (after ~17 shots in 4 star quality). In another test I could verify that a busy AF.C can indeed slow down the burst speed. Fortunately, I did run #5 over 50m/2sec and the K-7 still took 7 images :)

7 runs, per run (total=++/+/o/-/(subject out of frame):

1. 20=13/3/3/0/(1) too slow
2. 16=11/3/1/1 little too slow and little to far
3. 14=13/0/0/1 correct speed and distance!
4. 17=15/1/0/1 little too fast and too far
5. 7=5/0/1/0/(1) correct speed and 50m (run #5 is not counted!)
6. 20=19/0/0/1 correct speed and too far
7. 16=15/0/1/0 too slow and too short

Run #3 are 7% bad images only. And to quote a comparison: 13 good images is what the D300 delivered in the original test (rank 6/9)!

Fig.2: The 100% crop images for run #2 are all shown in the compound image above and these are the ratings: +/++(8x)/+/++(3x)/o/+/- (last one was too near to have any chance, and last one is not a crop).

In the same attempt to simplify my analysis, I get this result (counting out of frame as bad):

• 17.2=15.5/1.7 (9.9% bad; rank 4/9 -- about same as D300)

Summary and verdict:

The K-7 makes about 50% more good images and about 50% less bad images than the K20D in this test. This looks like a lot. But believe me: when I first saw the images, I thought ... "where the hell is the difference"? Both cameras for sure can capture driving cars except when they are very nearby. The difference is subtle more than evident. But on the other hand, it is significant enough to score side by side with a D300. Seems others also put their pants on one leg at a time. In practical terms, the AF.C will still fail in many situations when one would wish otherwise...

Additionally, the systematic errors in my test are large. So, let's me say it this way then:

The AF.C on the K-7, compared to the K20D, is an improvement making it significantly coming closer to the performance of a D300. How close exactly it comes, I don't know. Could be very close, or not.

This concludes PART I. The next part will study, in a just a bit less formal way, what happens with people running randomly around you ...

-> Continue to Part II