Can AF speed be improved via firmware?

[larbel]

Advertisement (gone after registration)

There're talk about the X1's AF speed is limited by the processor, but from my observation so far, I *think* there might be rooms to improve via firmware. With 1 point AF, it tends to run through the whole focus range most of the time before it locks, and with 1 point H, if the object is close from 1 frame to another, lots more often than 1 point, the X1 seeks out from the current focus position hence much faster lock. The same pretty much goes to 11 point and 11 point H.

 

On the other hand, with spot, it ALWAYS run though the entire ranges before locks no matter what. True that with the smaller meter, it has less contrast to detect, but it felt more like badly programmed focus which is very unefficient, I think at least it can be as fast as 1 point with some tweak via firmware, and there's definately rooms to make all the AF more efficient IMO... Thoughts anyone?

[seattlesteve]

It seems possible to increase the AF speed through a firmware upgrade, but I'm no engineer. I also read that AF speed was a function of the processor, but algorithms can always be improved. It's been done in the past. See this AF firmware upgrade from Sigma regarding the DP2: Firmware Ver.1.04 download notice for SIGMA DP2 - SIGMA CORPORATION

 

Here's hoping.

[Advertisement]

The Leica Q3 is available for pre-order at B&H Photo Video and Adorama

[mjh]

You have found that the high-speed (H) modes are indeed faster, and that selecting a smaller AF field results in longer focusing times. Just as one would have expected. But how does this suggest that the AF speed could be improved by a firmware update?

 

The AF is based on measuring contrast. Unfortunately a single contrast reading tells you nothing about focus – neither whether you have achieved optimum focus nor where (or even in which direction) the lens had to travel if you have not. If you wanted to be really thorough you would have to sample contrast readings at fine intervals along the whole focusing range, then travel back to the point where the highest contrast was measured. You could try to cut corners by starting at the current position, travel some small steps in an arbitrary direction, continuing in the same direction if the contrast readings improve and reversing direction if they get worse. You would continue in that direction until the readings get worse again, then return to the point with the highest reading. That would be faster than the more thorough approach but run the risk of getting stuck at a local maximum when a still higher contrast could have been found elsewhere. It is certainly possible that the algorithm implemented could be improved, but I don’t see any obvious indications of this. Especially since the algorithm was probably licensed from some other vendor (Panasonic being a likely candidate for more than one reason) with a wealth of experience in this area.

[eggfree]

So it depends on what's slow...

Incidentally I recall Sean Ried discussing this at some point, so you might ask him

 

I suspect (i.e. this is conjecture) the Leica can't readout fast enough, and can't focus fast enough (no ring motors, and no fast way to bin pixels or readout a subset of the sensor for low-light read performance. I remember seeing a deconstructed view of the lens assembly some while back and it appeared to have a conventional motor in body.

 

Contrast Detection AF relies on taking a series of fast sample images and comparing them.

+ Faster AF motors (coupled with a lens designed for fast autofocus)

+ Faster Sensor readout

+ ability to do partial sensor readout (if you are centerpoint af, no need to read all the MP)

+ Faster Processors

+ Optimal AF sampling (contrast detect uses samples imaged from a portion of the sensor captured at different points of focus to pick optimal focus, and there is a tradeoff in accuracy vs. speed, that is # samples)

 

The Panasonic is exceptionally fast supposedly because its sensor was designed for video, probably fast (on-chip?)pixel binning (video is 720p ) which implies 9 pixels sample into each evaluated pixel 60 times a second. The Leica may be about right compared to other contrast-detect point and shoots using a general purpose sensor.

the binning probably results in 1-1.5 stops of addtional sensitivity, and may reduce the post-binned compute load by a factor of 9, allowing for faster processing of contrast detect algorithms.

 

 

Since Sony will enter the big-sensor-little-camera market at some point, look for them to develop the appropriate sensor-hardware-firmware, then OEM it to others (which they have done in the past)

[mjh]

The Panasonic is exceptionally fast supposedly because its sensor was designed for video, probably fast (on-chip?)pixel binning (video is 720p ) which implies 9 pixels sample into each evaluated pixel 60 times a second. The Leica may be about right compared to other contrast-detect point and shoots using a general purpose sensor.

the binning probably results in 1-1.5 stops of addtional sensitivity, and may reduce the post-binned compute load by a factor of 9, allowing for faster processing of contrast detect algorithms.

You don’t use binning to kill fine contrast if you need to evaluate that contrast for determining optimum focus. There is no need to read out all the pixels (just the selected AF field), but you do need to use the same sensor resolution for focusing that you intend to use in the eventual image.

 

With a phase-detection autofocus, focusing speed crucially depends on a fast motor. The AF module can determine (nearly) optimum focus with just one measurement, so it just depends on how fast the motor can accurately hit the predetermined position. But contrast-based AF is quite different. The focus has to be changed in small increments, measuring contrast along the way. With small sensors this can be reasonably fast as there aren’t that many focus settings to consider; the depth-of-field that goes with a small sensor takes care of any small inaccuracies. But with a larger sensor the focus points multiply; everything else being equal the focusing time will increase with the sensor size.

[Dan States]

In one point AF the camera will at times focus almost immediately. This occurs when the point of focus is of a clearly offset color to it's surroundings and the entire object of focus is within the spot AF window. A moving object or an object with less defined shape will mean a longer focus cycle.

 

I can repeat this over and over again and at different distances. When it does focus this quickly the sound made is quite different from the hesitant sputtering it will typically do in many cases. It makes a fast "grunt" and focus it achieved.

 

To me this indicates that the lens is quite capable of focusing very quickly IF the AF spot can lock focus and that the AF spot CAN lock focus quickly, but only when conditions are ideal.

 

Considering how much faster I'm finding 11 pt AF to be in general photography Id say that the hardware is good enough and Leica should be working on the AF code and the size of the AF points.

[eggfree]

Advertisement (gone after registration)

You don’t use binning to kill fine contrast if you need to evaluate that contrast for determining optimum focus. There is no need to read out all the pixels (just the selected AF field), but you do need to use the same sensor resolution for focusing that you intend to use in the eventual image.

 

Ahh, you are presuming 3 pixels matter. Auto-focus algorithms may "leverage" the depth of field to get "close enough" to cover the circle of confusion, rather than nail focus at the highest pixel resolution.

 

In low light situations you have to contend with both a low contrast image close to the noise floor, and getting afast read from the sensor. That is where binning comes in handy. you potentially sacrifice some focus accuracy for a higher quality image to analyze (lower noise floor) with any luck, the selected f/stop will cover close enough.

 

The X1 team said they optimized for accuracy over speed, and it may relate directly to what shortcuts they didn't take.

 

take care,

JE

Edited April 8, 2010 by eggfree
clarification (meant low light, said high iso)

[mjh]

Ahh, you are presuming 3 pixels matter. Auto-focus algorithms may "leverage" the depth of field to get "close enough" to cover the circle of confusion, rather than nail focus at the highest pixel resolution.

In a pixel-peeping world, the permissible circle of confusion is defined by the sensor resolution; nothing less will do. So, yes, 3 pixels do matter.