Why is contrast AF considered to be inferior?

falconeye · 03-01-2008, 07:48 PM

I see that everybody seems to take it as granted that sensor AF as done by all SLRs including DSLRs is far better than "contrast AF" where the actual sensor data is used.

Can anybody provide arguments for this?

I cannot see a reason. I see that SLRs with their lack of a sensor had no other choice and early DSLRs probably just continueed to use what was already there. And then everybody noticed that "contrast AF" would darken the VF (mirror up), so not much research may have been invested in this direction. But none of this is an argument for one or the other method.

I've once implemented a "contrast AF" used on microscopes meant to replace the infrared laser AF commonplace in some of those systems.

Well, I ended up applying a smarter method than just using an image's contrast which I cannot disclose here. All I can say that it performed extremely fast, accurate and reliably with only a few "re-focus" steps involved (about 3-4). It easily met the hardware infrared laser AF specs.

I understand that photography may be different because lenses may be slower to re-focus and it may take much exposure time to take each snapshot in turn. On the other side, microscopes are harder to do as well because they have a DoF of a few micrometers only (which means manual focus is very hard while you see nothing more than gray soup...).

So, ideally supported by traditional focus sensors to predict the direction of re-focus, "contrast AF" should be faster, more reliably and far more accurate than the traditional method.

Anybody knows why it isn't done?

Last edited by falconeye; 03-01-2008 at 08:25 PM. Reason: Removed irrelevant detail

RBellavance · 03-02-2008, 05:24 AM

Contrast-detection AF is inherently slower, because it does not compute distance to the subject and must "search" until it finds the setting with greatest contrast.

Phase detection computes the distance to the focus point and moves the focus directly there (and then the Pentax implementation does its infamous "double-checks").

falconeye · 03-02-2008, 07:03 AM

Originally posted by RBellavance

Contrast-detection AF is inherently slower, because it does not compute distance to the subject and must "search" until it finds the setting with greatest contrast.

Phase detection computes the distance to the focus point and moves the focus directly there (and then the Pentax implementation does its infamous "double-checks").

While sounding correct, I fear this answer is given too quick by too many.

Sensor-based AF using phase cross correlation (like Safox) could be fast by pre-computing the off-focus distance. But the measurements across vendors says that it takes between 0.5s and 2s to focus. This is damned slow. The algorithm I did could focus within ms! Ok, microscopes have more light and the focus motors operated extremely quick to allow for a complete focus to complete within less than 0.1s. What I want to say is the following: a good "contrast AF" should not slow down the focussing at all; but make it more accurate and reliable.

A "contrast AF" can take intermediate measurements while the focus motor runs and if the initial direction of movement is predicted correctly by additional phase detection, does not cost time given good light. And at low light, only a deterministic extra amout of time (about 3x the exposure time, i.e., still <1s) would be required to gain good focus.

Still cannot see why it isn't done.

Meanwhile, I've read somewhere (cannot find it anymore, I googled for Safox) that Pentax is redoing their AF system in favour of a "sensor-image based tracking algorithm". So, maybe, they abondon phase detection as main method for AF and this is the reason why they only tweaked it a litlle bit for the K20D.

deejjjaaaa · 03-02-2008, 11:04 AM

Originally posted by falconeye

Sensor-based AF using phase cross correlation (like Safox) could be fast by pre-computing the off-focus distance. But the measurements across vendors says that it takes between 0.5s and 2s to focus. This is damned slow. The algorithm I did could focus within ms! Ok, microscopes have more light and the focus motors operated extremely quick to allow for a complete focus to complete within less than 0.1s.

what was the price on your "microscope" comparing w/ the camera/lenses ? and size ?

falconeye · 03-02-2008, 11:58 AM

Originally posted by deejjjaaaa

what was the price on your "microscope" comparing w/ the camera/lenses ? and size ?

This was expensive equipment. More expensive than cars if you know what I mean

But that's not the point. I know how much computing power was needed in my approach (not too much for current DSLRs) and that the focus motor has to shift focus by so much anyway and this was the slowest operation as it was mechanic.

Therefore, this isn't a matter of price/cost arguments at all.

ankit · 03-02-2008, 12:16 PM

One possible reason is that the image resolution in live view mode is much lower than the image resolution of the final static image. Any contrast based AF would work in the live-view mode (so that it is real-time), and can only achieve accuracy upto the live-view resolution. Phase based AF, however, can use arbitrarily higher res sensors (usually 1D CCDs, I think) to get much higher focus accuracy.

This may be why the Nikon D300 uses contrast based AF only when it has to, and phase based AF for situations that demand more accuracy.

deejjjaaaa · 03-02-2008, 04:43 PM

Originally posted by falconeye

This was expensive equipment. More expensive than cars if you know what I mean

But that's not the point.

no that is the point - for that kind of money you can get a real fast focusing hardware - I mean
the hardware that actually moves/rotates those lenses after receiving commands from AF system... so that it will not be "...0.5s and 2s.." but "...within ms..."

so please compare things with $$$ in mind - apples to apples.

Gooshin · 03-02-2008, 05:13 PM

Originally posted by deejjjaaaa

no that is the point - for that kind of money you can get a real fast focusing hardware - I mean
the hardware that actually moves/rotates those lenses after receiving commands from AF system... so that it will not be "...0.5s and 2s.." but "...within ms..."

so please compare things with $$$ in mind - apples to apples.

our "hardware" is already fast, all of the new lenses with HSM/USM/SDM whatever are all lightning quick

pentax lenses focus mechanicaly as quick as any nikon or canon lens, its the software that tells it where and how to focus that sucks

ie, software

falconeye · 03-04-2008, 05:21 AM

Originally posted by Gooshin

ie, software

Thank You so much for clarifying.

As for the limited live view resolution. I did not propose a system which focusses like some of the other DSLRs in LiveView mode. I proposed a system where, while the camera is running its focus motor, full scale images are taken and evaluated into a focus metering (can be done in less than 100ms each, I did it). Yes, with current mirrors, the VF would dark out for a second while focussing. However, an easy invention then would be a mirror letting 1/8 of light pass permanently to the sensor (so only at low light would the VF dark out -- and LiveView would not always dark out the VF as well).

03-01-2008, 07:48 PM	#1
falconeye Veteran Member Join Date: Jan 2008 Location: Munich, Alps, Germany Photos: Gallery Posts: 6,871	Why is contrast AF considered to be inferior? I see that everybody seems to take it as granted that sensor AF as done by all SLRs including DSLRs is far better than "contrast AF" where the actual sensor data is used. Can anybody provide arguments for this? I cannot see a reason. I see that SLRs with their lack of a sensor had no other choice and early DSLRs probably just continueed to use what was already there. And then everybody noticed that "contrast AF" would darken the VF (mirror up), so not much research may have been invested in this direction. But none of this is an argument for one or the other method. I've once implemented a "contrast AF" used on microscopes meant to replace the infrared laser AF commonplace in some of those systems. Well, I ended up applying a smarter method than just using an image's contrast which I cannot disclose here. All I can say that it performed extremely fast, accurate and reliably with only a few "re-focus" steps involved (about 3-4). It easily met the hardware infrared laser AF specs. I understand that photography may be different because lenses may be slower to re-focus and it may take much exposure time to take each snapshot in turn. On the other side, microscopes are harder to do as well because they have a DoF of a few micrometers only (which means manual focus is very hard while you see nothing more than gray soup...). So, ideally supported by traditional focus sensors to predict the direction of re-focus, "contrast AF" should be faster, more reliably and far more accurate than the traditional method. Anybody knows why it isn't done? Last edited by falconeye; 03-01-2008 at 08:25 PM. Reason: Removed irrelevant detail

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
The Cheap, Nasty, Inferior & Underrated Mother Of A Lens Club	Christopher M.W.T	Lens Clubs	41	02-27-2019 01:33 PM
what's considered fast enough?	Deiberson	Troubleshooting and Beginner Help	19	01-16-2010 10:49 AM
Why are the K20D and K7 not considered Pro	fccwpe	Pentax DSLR Discussion	23	12-11-2009 05:09 PM
What is considered correct expsure?	schmik	Photographic Technique	34	08-28-2009 01:25 PM
How many shots is considered to much for a used camera?	Sitting Bull	Troubleshooting and Beginner Help	8	02-02-2009 01:40 PM

03-02-2008, 05:24 AM	#2
RBellavance Veteran Member Join Date: Nov 2006 Location: Near Montréal, Canada Photos: Gallery Posts: 1,716	Contrast-detection AF is inherently slower, because it does not compute distance to the subject and must "search" until it finds the setting with greatest contrast. Phase detection computes the distance to the focus point and moves the focus directly there (and then the Pentax implementation does its infamous "double-checks").

03-02-2008, 07:03 AM	#3
falconeye Veteran Member Join Date: Jan 2008 Location: Munich, Alps, Germany Photos: Gallery Posts: 6,871 Original Poster	Originally posted by RBellavance Contrast-detection AF is inherently slower, because it does not compute distance to the subject and must "search" until it finds the setting with greatest contrast. Phase detection computes the distance to the focus point and moves the focus directly there (and then the Pentax implementation does its infamous "double-checks"). While sounding correct, I fear this answer is given too quick by too many. Sensor-based AF using phase cross correlation (like Safox) could be fast by pre-computing the off-focus distance. But the measurements across vendors says that it takes between 0.5s and 2s to focus. This is damned slow. The algorithm I did could focus within ms! Ok, microscopes have more light and the focus motors operated extremely quick to allow for a complete focus to complete within less than 0.1s. What I want to say is the following: a good "contrast AF" should not slow down the focussing at all; but make it more accurate and reliable. A "contrast AF" can take intermediate measurements while the focus motor runs and if the initial direction of movement is predicted correctly by additional phase detection, does not cost time given good light. And at low light, only a deterministic extra amout of time (about 3x the exposure time, i.e., still <1s) would be required to gain good focus. Still cannot see why it isn't done. Meanwhile, I've read somewhere (cannot find it anymore, I googled for Safox) that Pentax is redoing their AF system in favour of a "sensor-image based tracking algorithm". So, maybe, they abondon phase detection as main method for AF and this is the reason why they only tweaked it a litlle bit for the K20D.

03-02-2008, 11:04 AM	#4
deejjjaaaa Veteran Member Join Date: May 2007 Location: steel city / rust belt Posts: 2,043	Originally posted by falconeye Sensor-based AF using phase cross correlation (like Safox) could be fast by pre-computing the off-focus distance. But the measurements across vendors says that it takes between 0.5s and 2s to focus. This is damned slow. The algorithm I did could focus within ms! Ok, microscopes have more light and the focus motors operated extremely quick to allow for a complete focus to complete within less than 0.1s. what was the price on your "microscope" comparing w/ the camera/lenses ? and size ?

03-02-2008, 11:58 AM	#5
falconeye Veteran Member Join Date: Jan 2008 Location: Munich, Alps, Germany Photos: Gallery Posts: 6,871 Original Poster	Originally posted by deejjjaaaa what was the price on your "microscope" comparing w/ the camera/lenses ? and size ? This was expensive equipment. More expensive than cars if you know what I mean But that's not the point. I know how much computing power was needed in my approach (not too much for current DSLRs) and that the focus motor has to shift focus by so much anyway and this was the slowest operation as it was mechanic. Therefore, this isn't a matter of price/cost arguments at all.

03-02-2008, 12:16 PM	#6
ankit Senior Member Join Date: Dec 2007 Location: Cambridge, MA Posts: 140	One possible reason is that the image resolution in live view mode is much lower than the image resolution of the final static image. Any contrast based AF would work in the live-view mode (so that it is real-time), and can only achieve accuracy upto the live-view resolution. Phase based AF, however, can use arbitrarily higher res sensors (usually 1D CCDs, I think) to get much higher focus accuracy. This may be why the Nikon D300 uses contrast based AF only when it has to, and phase based AF for situations that demand more accuracy.

03-02-2008, 04:43 PM	#7
deejjjaaaa Veteran Member Join Date: May 2007 Location: steel city / rust belt Posts: 2,043	Originally posted by falconeye This was expensive equipment. More expensive than cars if you know what I mean But that's not the point. no that is the point - for that kind of money you can get a real fast focusing hardware - I mean the hardware that actually moves/rotates those lenses after receiving commands from AF system... so that it will not be "...0.5s and 2s.." but "...within ms..." so please compare things with $$$ in mind - apples to apples.

03-02-2008, 05:13 PM	#8
Gooshin Veteran Member Join Date: Sep 2007 Location: Toronto, the one in Canada. Posts: 5,610	Originally posted by deejjjaaaa no that is the point - for that kind of money you can get a real fast focusing hardware - I mean the hardware that actually moves/rotates those lenses after receiving commands from AF system... so that it will not be "...0.5s and 2s.." but "...within ms..." so please compare things with $$$ in mind - apples to apples. our "hardware" is already fast, all of the new lenses with HSM/USM/SDM whatever are all lightning quick pentax lenses focus mechanicaly as quick as any nikon or canon lens, its the software that tells it where and how to focus that sucks ie, software

03-04-2008, 05:21 AM	#9
falconeye Veteran Member Join Date: Jan 2008 Location: Munich, Alps, Germany Photos: Gallery Posts: 6,871 Original Poster	Originally posted by Gooshin ie, software Thank You so much for clarifying. As for the limited live view resolution. I did not propose a system which focusses like some of the other DSLRs in LiveView mode. I proposed a system where, while the camera is running its focus motor, full scale images are taken and evaluated into a focus metering (can be done in less than 100ms each, I did it). Yes, with current mirrors, the VF would dark out for a second while focussing. However, an easy invention then would be a mirror letting 1/8 of light pass permanently to the sensor (so only at low light would the VF dark out -- and LiveView would not always dark out the VF as well).