Two statistics majors went on a police ride-along to see how the new crime prediction model was working. They went to the first predicted location on 200th st., but there was no crime. Then they went to the second second predicted location on 100th st, but again no crime. Just when they were about to admit defeat, a call came in about a crime on 150th st and they both yelled "we did it!"
It's amazing how you think you can extrapolate a doctor's salary from 5 minutes.
It doesn't matter whether the $450 for a 5-minute procedure goes to the doctor's salary or to clinic's investors. That's not the point.
Also, you underestimate the damage of losing the first four earning years out of college and instead accruing a house's worth of loans. Not to mention at least another 3-4 years after that where you're making a resident's salary.
Are you saying that $450 is justified because of that?
... the strongest concoction you get from her is fucking water you moron.
Thanks for that, friend, but I haven't gotten anything from her except liquid nitrogen and a variety of prescriptions written on paper. Do homeopaths have a special deal with pharmacies and billion-dollar pharmaceutical corporations where my prescriptions for Zyrtec, Nasonex, Proventil, etc. get secretly switched with water if they are written by a homeopath? Because they sure seemed identical to the drugs I got with the MD's prescription in the previous year.
Thanks. I wonder why the Empire Blue fraud department said those were the correct codes for a 5-minute wart spray visit with the PCP. But maybe they were just giving me the quick brush off because they didn't want to bother with the hassle over a measly $450.
The fact that your employer pays $1700 a month for a Cadillac health care plan for you completely baffles me. Why would they spend that much money per employee on straight health care?
I don't really know, but as I hinted in the original post, my guess is that it's due to our braindamaged tax laws:
* Employer plans are tax deductible, while individual plans are not. In my tax bracket, that means I can get a $1700/mo plan for the same effective cost as a $1100 plan if I were to buy it using increased income and an individual plan. If individual premiums were deductible above the line, I'd rather take the $1700/mo in salary and buy whatever plan I actually needed. * Avoiding the painfully retarded double taxation (dividend tax) by compensating owners through insurance. * IRS Employee "equality" rules forbidding owners from having nicer insurance than employees. * Avoiding non-deductible health care costs. For example, paying an extra $500/mo in deductible premiums instead of $500/month in premiums, copays, co-insurance (which still wouldn't be enough to reach the 7.5% rule).
That's just fucking crazy, god i'm happy that i live in a sane country. Over here we pay something like 40$ for "healthcare tax" monthly from your salary
I wonder. What percent of the *actual* cost does $40/month cover? What pays for the rest of it? And what would a 5-minute wart spray cost in your country?
In short, $450 for a quick realistic medical procedure, while wildly overpriced, is still not nearly as overpriced as $40 "water with memory" treatment
Huh? She used the exact same liquid nitrogen and technique as the medical doctor. The only difference is several orders of magnitude in the price. Oh, and also that I didn't have to call ahead and wait several weeks for an appointment, nor did I have to sit for an hour in the waiting room and another 20 minutes in the examination room.
Their profits decrease as costs increase, and they do care about minimizing costs.
By what mechanism, exactly? Is it due to competition? For example, if costs increase, the insurance company could retain the same profit margin by just increasing premiums. Subscribers will either pay the higher premium or drop down to a lower-benefit plan. But if a competing health insurance company achieves lower cost (by paying less than $5,400/hour, for example), then subscribers may switch to the competition (which can provide the same benefit at a lower premium). Is that the mechanism?
You may be right about the application of Hanlon's razor in this case.
The first sentence of the summary says they "spotted" the brown dwarf. This implies that it was out there and they observed it.
You are mistaken. When they "spotted" the brown dwarf, it means they added decorative spots to it. So not only do they have the power to control its temperature as you correctly pointed out, but now they can even add polka dots. I'm looking forward to plaid stars.
It's not going to take 30 years for that system to fit in your pocket and cost $20. It's going to take 5 or 10.
No. It will be far more than 10 years before you buy 15 TB RAM and 2880 cores for $20, let alone fit it in your pocket.
Even 30 years is pushing it. Let's examine just the CPU. Watson has a total of 432 billion transistors in 2.2 square *feet* total die area (204120mm^2) using 45nm process. By contrast, the iPhone 4 has one core in 53mm^2 using 45nm process.
That's *four* orders of magnitude. When I do the math, I come up with 48 years for transistors to shrink that much (see below). That's assuming Moore's Law holds true indefinitely and $20 portables will have 50mm^2 processors.
The only way I can get 10 years is if I assume that $20 portable computers will have 35,000mm^2 die area for the CPU. That's a 7-inch-wide CPU die (!). By contrast, the Intel Core i5 die is only half an inch in diameter.
At best, we'll get enough software optimizations that we can run something "similar" to Watson on much less powerful hardware, such as $20 portables.
The math (works in octave but not bc):
# IBM power7 die area in square mm. per_cpu_da = 567
# Number of power7 processors in Watson watson_cpu_count = 360
# Watson die area in square mm. watson_da = per_cpu_da * watson_cpu_count
# Cheap and portable die area in square mm. cheap_da = 50
# Number of years for transistor count to double for a given area. doubles_every = 2
Most fires result in near or total destruction even with the intervention of dozens of firefighters with millions of dollars of equipment.
Do you have any references for that? Among my family and acquaintances there have been two house fires with the intervention of dozens of firefighters with millions of dollars of equipment, and neither one resulted in near or total destruction. Just a refinished garage and bathroom.
you can only ever get a 50% MTF at around 250lp/mm and f2.8. For 80% MTF in the same resolution, you need about f1.2.
Excellent point. Those are the same numbers I get.
As I'm sure you know, a lot of contrast can be restored with sharpening or deconvolution. 50% MTF and even as low as 20% MTF can be restored to pretty high contrast. But it only works well if noise is low enough.
To me, it would be ideal to have a pixel size that results in less than 15% MTF for typical f-numbers (say, f/5.6, which puts us at about the same pixel size Canon used here). That way we could finally get rid of the anti-alias filter and improve MTF noticeably at a lot of lower spatial frequencies without aliasing artifacts.
So at the majority of working apertures, you're probably not going to be able to tell between a photo shot with a 120Mpixel sensor and one scaled up from an 60Mpixel sensor on a 35mm-class camera... and that's with a very good lens.
Agreed. Not to mention the much more common case where resolution is diminished even more by missed focus, motion blur, lens aberration, poor collimation, etc.
Yet despite all these reasons why most circumstances wont benefit from such small pixels, I would still like to see it happen. I think the downsides are small for most common use cases (ample light, normal dynamic range). The benefit will be for those rare but useful times when the resolution does come in handy, such as making teleconverters obsolete.
One of the most commonly cited downsides is file size. But for me, RED ONE has already proven that this is a solved problem: with the right software and sufficient encoding power in the camera, we can keep the overall quality to file size ratio the same at any given pixel count. So a 120 MP heavily compressed to 10 MB will give the same quality as a 20 MP compressed to 10 MB. Of course, in that situation the higher pixel count would not offer a benefit -- that is not the intention -- but it would at least allow for the pixel count to not be a disadvantage. Then the same pixel count can be used for lots of purposes. Those who have the space for large files can use lower compression or none at all, while those would prefer to keep the same file size as lower pixel count sensors can do that too.
Another oft-discussed disadvantage is processing speed, such as for raw conversion, but I think this too can be scaled linearly with overall image quality, at least for powers of two scaling factors, as demonstrated by the alternate demosaic in Raw Photo Processor. Higher quality solutions may be possible in the same time if execution is sufficiently parallelized and Moore's Law stays ahead of pixel count increases.
Compare this to a Canon 5D Mk II sensor... the pixel pitch on this model is 6.4mm x 6.4mm. That's nearly 8.5x the area per pixel.
Yes, a single 6.4 micron pixel has much higher sensitivity than a single 2.2 micron pixel, but that's not seeing the forest for the trees. The higher number of pixels perfectly compensates for the lower sensitivity in the same way that a eight small cylinders can achieve the same total horsepower as four large cylinders if the displacement is the same for both engines. It's not the horsepower-per-cylinder that matters, it's the total horsepower of the engine.
The reason for this equivalence as it pertains to sensitivity is that the noise power at any given spatial frequency (level of detail) is equal as long as the sensitivity per square micron is equal.
Smaller pixels sample detail at a higher spatial frequency. If the noise power at that higher level of detail is compared directly to larger pixels which sample lower frequencies, it will be concluded the small pixel sensor is worse overall when in fact it is not, just as measuring hp-per-cylinder would lead to the incorrect conclusion that a small engine with large cylinders is better than a large engine with small cylinders. In reality, the small pixel sensor achieves the same noise power as the large pixel sensor for any given spatial frequency, it just has the added and optional capability to sample higher spatial frequencies.
The 5D2 you brought up is an excellent illustration of this. Despite having much larger pixels than the 7D, its noise power for any given level of detail and amount of sensor area is the exact same.
(Since we're only talking about sensitivity and not read noise, all this of course applies only to circumstances where the system noise is dominated photon shot noise.)
All a micro-lens can do for you is make a smaller sensor seem to occupy the full sensor-cell space available to it.
Right, and that is precisely what allows it to have the same sensitivity (photoelectron carriers per square micron), whereas otherwise a larger portion of the pixel area would have been taken by non-photodiode elements (for a given minimum feature size). Without microlenses, not only would smaller pixels have lower sensitivity per pixel, but they would also have lower sensitivity per area (e-/um^2).
As long as the rate of decline in full-well capacity matches the rate of increase in spatial-frequency-normalized read noise on the slope of pixel size, the dynamic range will stay the same, not decrease. In practice, The rate of decline is usually slower, particularly if you restrict the dynamic range to the circumstance of a typical photographer, who doesn't care for SNR lower than 8:1, and therefore will be photon-shot-noise dominated, leaving the FWC, not read noise, as the dominant factor.
That why dynamic range is so good in small pixels (if you factor out the overall sensor size, of course), like 10.3 stops in the cheap G10 at SNR 1:1, which slightly more then the $2,500 Canon 5D2 but not as good as the 12-stop D3X for $8,000.
The point is this: 120 MP allows for a choice between resolution and dynamic range. One the one hand, full resolution and normal dynamic range (e.g. 10 stops at 1:1 SNR), or less resolution and more dynamic range (e.g. 30 MP+11 stops, or 8 MP and 12 stops).
Unfortunately, this is not the case for low light performance, because it is dominated by read noise instead of photon shot noise. Unless a method is found to scale read noise linearly with pixel size, larger pixels will generally have an advantage over small pixels even after normalizing for spatial frequency.
This is a solved problem. RED ONE has proven that software can scale file size linearly with quality so that there is no file-size or computing speed disadvantage to larger pixel counts. (Not to mention that the typical balance between quality and size is almost an order of magnitude worse than what it could be.) Now if only the rest of the industry would start paying attention to the software engineers...
Ken Rockwell is to photography what a goatse troll is to Slashdot. (In fact, if you read his alien abduction pages, you'll see some similarity with goatse).
It's like saying "Computer specs don't matter. Unless you're folding proteins, a 486 is just as good as i5." While it's true that sharpness and resolution are not the most important factors in a photograph, it's misleading as their benefits do in fact contribute to most styles of photography, just as a faster computer can contribute to a better experience for most computing needs.
For example, most people feel that for an 8x10, there is no benefit to pixel counts above 6 MP, but in fact it takes a 24 MP before all the possible gains are realized, most importantly counteracting the loss in contrast from the anti-alias filter. (Many more MP would be required to hit full color resolution at Nyquist, but few natural images benefit from that, despite what the Foveon advocates claim.)
That's right--even "fast" f/2.8 lenses are limited by physics to never being able to project detail onto individual pixels.
That is incorrect. Parity between the airy disk and pixel diameter is not the point at which additional detail becomes impossible -- that is only one point on the curve of diminishing returns. In other words, it is the difference between the "diffraction limited" spatial frequency and the "diffraction cutoff" spatial frequency. It is only the latter that denotes the impossibility of further resolution from decreased pixel size.
The easiest way to understand this is to look at MTF. When diffraction causes the optical system MTF to drop to 50%, most would consider that the end of the line. But in fact, that is just the point where a lot of contrast is lost -- detail is still there and contrast can be restored with sharpening (e.g. RL deconvolution). MTF must drop to 10% before detail truly becomes extinct, and for a 2.2 micron pixel like this 120 MP Canon, f/5.6 will still give you 18% MTF, and there are a host of lenses that are very sharp at f/5.6.
For further consideration, look at the effect of the anti-alias filter, which drops MTF of spatial frequences far lower than needed to suppress aliasing. The ideal solution to this problem is pixels that are so small that diffraction itself anti-aliases. That will increase contrast at lower spatial frequencies by 30%.
Instead of worrying about "wasted pixels", we should be worrying about "wasted glass". A simple prime lens has about the same resolution and cost today as it did 20 years ago, whereas image sensor resolution and cost have advanced by many orders of magnitude in the same amount of time. Therefore we should be more concerned about that part of the system which is expensive and difficult to improve; not the part that is getting cheaper and better.
It's true that some lenses are already into diminishing returns, particularly fast or wide lenses. But many other lenses are not even oversampled with 2 micron pixel sizes, including many macro and sharp primes. Take a look at this example of a $400 macro lens using 1.2 micron (simulated) pixel size:
Dynamic range is the distance between clipping and noise. The standard engineering definition assumes a SNR of 1:1 as the lower bound, but few photographers can tolerate that much noise and usually prefer 4:1 or 8:1. Random noise sources add in quadrature so that downsampling the pixel count by a factor of four increases the SNR by a factor of 2. A better way of thinking about it is this: the raw data from an image sensor has a noise power that increases linearly with spatial frequency (level of detail). Higher frequency (smaller details) have higher noise powers. If you throw away the high frequency detail, the noise goes with it. In actual practice, there are many better ways to reduce noise than by throwing away detail, and in any case, many viewers will prefer a detailed but noisey image over a blurry but less noisy one.
Actually, the "sensitivity" (more specifically, e-/m^2) is generally the same across a huge variety of pixel sizes, thanks to microlenses. What is not usually constant is read noise (AKA "high ISO noise", sometimes also referred to as "sensitivity"), because although it does naturally shrink a little bit as the pixel size is reduced, it's not always in exact linear proportion with pixel diameter, and therefore the generalization that smaller pixels tend to have slightly more noise in low light.
Assume the car approaching from behind will not stop. You probably could have avoided that accident.
Several years ago, I was idling at a red light in Idaho during the Christmas shopping season. Traffic was greater than usual, so the line at the stop light had 20+ cars, much more than normal. Screeching tires and a split-second later the back-end of my Camry was smushed to bits. The driver had just barely touched their brakes, but I thought that I could have avoided it if I had been paying more attention.
For a while I was careful to notice any vehicle that approached quickly. Shortly I realized that it was futile. Too many drivers race to their position and slam on the brakes at the last second.
It's impossible to detect danger when so many drivers incorporate it as a regular part of their driving behavior.
Slashdot Mirror
Two statistics majors went on a police ride-along to see how the new crime prediction model was working. They went to the first predicted location on 200th st., but there was no crime. Then they went to the second second predicted location on 100th st, but again no crime. Just when they were about to admit defeat, a call came in about a crime on 150th st and they both yelled "we did it!"
It's amazing how you think you can extrapolate a doctor's salary from 5 minutes.
It doesn't matter whether the $450 for a 5-minute procedure goes to the doctor's salary or to clinic's investors. That's not the point.
Also, you underestimate the damage of losing the first four earning years out of college and instead accruing a house's worth of loans. Not to mention at least another 3-4 years after that where you're making a resident's salary.
Are you saying that $450 is justified because of that?
... the strongest concoction you get from her is fucking water you moron.
Thanks for that, friend, but I haven't gotten anything from her except liquid nitrogen and a variety of prescriptions written on paper. Do homeopaths have a special deal with pharmacies and billion-dollar pharmaceutical corporations where my prescriptions for Zyrtec, Nasonex, Proventil, etc. get secretly switched with water if they are written by a homeopath? Because they sure seemed identical to the drugs I got with the MD's prescription in the previous year.
Based on your description this was billed wrong.
Thanks. I wonder why the Empire Blue fraud department said those were the correct codes for a 5-minute wart spray visit with the PCP. But maybe they were just giving me the quick brush off because they didn't want to bother with the hassle over a measly $450.
The fact that your employer pays $1700 a month for a Cadillac health care plan for you completely baffles me. Why would they spend that much money per employee on straight health care?
I don't really know, but as I hinted in the original post, my guess is that it's due to our braindamaged tax laws:
* Employer plans are tax deductible, while individual plans are not. In my tax bracket, that means I can get a $1700/mo plan for the same effective cost as a $1100 plan if I were to buy it using increased income and an individual plan. If individual premiums were deductible above the line, I'd rather take the $1700/mo in salary and buy whatever plan I actually needed.
* Avoiding the painfully retarded double taxation (dividend tax) by compensating owners through insurance.
* IRS Employee "equality" rules forbidding owners from having nicer insurance than employees.
* Avoiding non-deductible health care costs. For example, paying an extra $500/mo in deductible premiums instead of $500/month in premiums, copays, co-insurance (which still wouldn't be enough to reach the 7.5% rule).
That's just fucking crazy, god i'm happy that i live in a sane country. Over here we pay something like 40$ for "healthcare tax" monthly from your salary
I wonder. What percent of the *actual* cost does $40/month cover? What pays for the rest of it? And what would a 5-minute wart spray cost in your country?
In short, $450 for a quick realistic medical procedure, while wildly overpriced, is still not nearly as overpriced as $40 "water with memory" treatment
Huh? She used the exact same liquid nitrogen and technique as the medical doctor. The only difference is several orders of magnitude in the price. Oh, and also that I didn't have to call ahead and wait several weeks for an appointment, nor did I have to sit for an hour in the waiting room and another 20 minutes in the examination room.
Their profits decrease as costs increase, and they do care about minimizing costs.
By what mechanism, exactly? Is it due to competition? For example, if costs increase, the insurance company could retain the same profit margin by just increasing premiums. Subscribers will either pay the higher premium or drop down to a lower-benefit plan. But if a competing health insurance company achieves lower cost (by paying less than $5,400/hour, for example), then subscribers may switch to the competition (which can provide the same benefit at a lower premium). Is that the mechanism?
You may be right about the application of Hanlon's razor in this case.
The first sentence of the summary says they "spotted" the brown dwarf. This implies that it was out there and they observed it.
You are mistaken. When they "spotted" the brown dwarf, it means they added decorative spots to it. So not only do they have the power to control its temperature as you correctly pointed out, but now they can even add polka dots. I'm looking forward to plaid stars.
It's not going to take 30 years for that system to fit in your pocket and cost $20. It's going to take 5 or 10.
No. It will be far more than 10 years before you buy 15 TB RAM and 2880 cores for $20, let alone fit it in your pocket.
Even 30 years is pushing it. Let's examine just the CPU. Watson has a total of 432 billion transistors in 2.2 square *feet* total die area (204120mm^2) using 45nm process. By contrast, the iPhone 4 has one core in 53mm^2 using 45nm process.
That's *four* orders of magnitude. When I do the math, I come up with 48 years for transistors to shrink that much (see below). That's assuming Moore's Law holds true indefinitely and $20 portables will have 50mm^2 processors.
The only way I can get 10 years is if I assume that $20 portable computers will have 35,000mm^2 die area for the CPU. That's a 7-inch-wide CPU die (!). By contrast, the Intel Core i5 die is only half an inch in diameter.
At best, we'll get enough software optimizations that we can run something "similar" to Watson on much less powerful hardware, such as $20 portables.
The math (works in octave but not bc):
# IBM power7 die area in square mm.
per_cpu_da = 567
# Number of power7 processors in Watson
watson_cpu_count = 360
# Watson die area in square mm.
watson_da = per_cpu_da * watson_cpu_count
# Cheap and portable die area in square mm.
cheap_da = 50
# Number of years for transistor count to double for a given area.
doubles_every = 2
doubling_factor = 1 / doubles_every
years = log(cheap_da / watson_da) / (doubling_factor * log(1 / sqrt(2)))
Most fires result in near or total destruction even with the intervention of dozens of firefighters with millions of dollars of equipment.
Do you have any references for that? Among my family and acquaintances there have been two house fires with the intervention of dozens of firefighters with millions of dollars of equipment, and neither one resulted in near or total destruction. Just a refinished garage and bathroom.
Thank you for the great post, Dave.
you can only ever get a 50% MTF at around 250lp/mm and f2.8. For 80% MTF in the same resolution, you need about f1.2.
Excellent point. Those are the same numbers I get.
As I'm sure you know, a lot of contrast can be restored with sharpening or deconvolution. 50% MTF and even as low as 20% MTF can be restored to pretty high contrast. But it only works well if noise is low enough.
To me, it would be ideal to have a pixel size that results in less than 15% MTF for typical f-numbers (say, f/5.6, which puts us at about the same pixel size Canon used here). That way we could finally get rid of the anti-alias filter and improve MTF noticeably at a lot of lower spatial frequencies without aliasing artifacts.
So at the majority of working apertures, you're probably not going to be able to tell between a photo shot with a 120Mpixel sensor and one scaled up from an 60Mpixel sensor on a 35mm-class camera... and that's with a very good lens.
Agreed. Not to mention the much more common case where resolution is diminished even more by missed focus, motion blur, lens aberration, poor collimation, etc.
Yet despite all these reasons why most circumstances wont benefit from such small pixels, I would still like to see it happen. I think the downsides are small for most common use cases (ample light, normal dynamic range). The benefit will be for those rare but useful times when the resolution does come in handy, such as making teleconverters obsolete.
One of the most commonly cited downsides is file size. But for me, RED ONE has already proven that this is a solved problem: with the right software and sufficient encoding power in the camera, we can keep the overall quality to file size ratio the same at any given pixel count. So a 120 MP heavily compressed to 10 MB will give the same quality as a 20 MP compressed to 10 MB. Of course, in that situation the higher pixel count would not offer a benefit -- that is not the intention -- but it would at least allow for the pixel count to not be a disadvantage. Then the same pixel count can be used for lots of purposes. Those who have the space for large files can use lower compression or none at all, while those would prefer to keep the same file size as lower pixel count sensors can do that too.
Another oft-discussed disadvantage is processing speed, such as for raw conversion, but I think this too can be scaled linearly with overall image quality, at least for powers of two scaling factors, as demonstrated by the alternate demosaic in Raw Photo Processor. Higher quality solutions may be possible in the same time if execution is sufficiently parallelized and Moore's Law stays ahead of pixel count increases.
Compare this to a Canon 5D Mk II sensor... the pixel pitch on this model is 6.4mm x 6.4mm. That's nearly 8.5x the area per pixel.
Yes, a single 6.4 micron pixel has much higher sensitivity than a single 2.2 micron pixel, but that's not seeing the forest for the trees. The higher number of pixels perfectly compensates for the lower sensitivity in the same way that a eight small cylinders can achieve the same total horsepower as four large cylinders if the displacement is the same for both engines. It's not the horsepower-per-cylinder that matters, it's the total horsepower of the engine.
The reason for this equivalence as it pertains to sensitivity is that the noise power at any given spatial frequency (level of detail) is equal as long as the sensitivity per square micron is equal.
Smaller pixels sample detail at a higher spatial frequency. If the noise power at that higher level of detail is compared directly to larger pixels which sample lower frequencies, it will be concluded the small pixel sensor is worse overall when in fact it is not, just as measuring hp-per-cylinder would lead to the incorrect conclusion that a small engine with large cylinders is better than a large engine with small cylinders. In reality, the small pixel sensor achieves the same noise power as the large pixel sensor for any given spatial frequency, it just has the added and optional capability to sample higher spatial frequencies.
The 5D2 you brought up is an excellent illustration of this. Despite having much larger pixels than the 7D, its noise power for any given level of detail and amount of sensor area is the exact same.
(Since we're only talking about sensitivity and not read noise, all this of course applies only to circumstances where the system noise is dominated photon shot noise.)
All a micro-lens can do for you is make a smaller sensor seem to occupy the full sensor-cell space available to it.
Right, and that is precisely what allows it to have the same sensitivity (photoelectron carriers per square micron), whereas otherwise a larger portion of the pixel area would have been taken by non-photodiode elements (for a given minimum feature size). Without microlenses, not only would smaller pixels have lower sensitivity per pixel, but they would also have lower sensitivity per area (e-/um^2).
As long as the rate of decline in full-well capacity matches the rate of increase in spatial-frequency-normalized read noise on the slope of pixel size, the dynamic range will stay the same, not decrease. In practice, The rate of decline is usually slower, particularly if you restrict the dynamic range to the circumstance of a typical photographer, who doesn't care for SNR lower than 8:1, and therefore will be photon-shot-noise dominated, leaving the FWC, not read noise, as the dominant factor.
That why dynamic range is so good in small pixels (if you factor out the overall sensor size, of course), like 10.3 stops in the cheap G10 at SNR 1:1, which slightly more then the $2,500 Canon 5D2 but not as good as the 12-stop D3X for $8,000.
The point is this: 120 MP allows for a choice between resolution and dynamic range. One the one hand, full resolution and normal dynamic range (e.g. 10 stops at 1:1 SNR), or less resolution and more dynamic range (e.g. 30 MP+11 stops, or 8 MP and 12 stops).
Unfortunately, this is not the case for low light performance, because it is dominated by read noise instead of photon shot noise. Unless a method is found to scale read noise linearly with pixel size, larger pixels will generally have an advantage over small pixels even after normalizing for spatial frequency.
I wish they would spend more time on improving the dynamic range than to just play the megapixel count wars.
Agreed. The pattern noise in their $2,500 5D2 limits the dynamic range to even lower than their cheap little G11. It's embarrassing.
This is a solved problem. RED ONE has proven that software can scale file size linearly with quality so that there is no file-size or computing speed disadvantage to larger pixel counts. (Not to mention that the typical balance between quality and size is almost an order of magnitude worse than what it could be.) Now if only the rest of the industry would start paying attention to the software engineers...
Ken Rockwell is to photography what a goatse troll is to Slashdot. (In fact, if you read his alien abduction pages, you'll see some similarity with goatse).
It's like saying "Computer specs don't matter. Unless you're folding proteins, a 486 is just as good as i5." While it's true that sharpness and resolution are not the most important factors in a photograph, it's misleading as their benefits do in fact contribute to most styles of photography, just as a faster computer can contribute to a better experience for most computing needs.
For example, most people feel that for an 8x10, there is no benefit to pixel counts above 6 MP, but in fact it takes a 24 MP before all the possible gains are realized, most importantly counteracting the loss in contrast from the anti-alias filter. (Many more MP would be required to hit full color resolution at Nyquist, but few natural images benefit from that, despite what the Foveon advocates claim.)
That's right--even "fast" f/2.8 lenses are limited by physics to never being able to project detail onto individual pixels.
That is incorrect. Parity between the airy disk and pixel diameter is not the point at which additional detail becomes impossible -- that is only one point on the curve of diminishing returns. In other words, it is the difference between the "diffraction limited" spatial frequency and the "diffraction cutoff" spatial frequency. It is only the latter that denotes the impossibility of further resolution from decreased pixel size.
The easiest way to understand this is to look at MTF. When diffraction causes the optical system MTF to drop to 50%, most would consider that the end of the line. But in fact, that is just the point where a lot of contrast is lost -- detail is still there and contrast can be restored with sharpening (e.g. RL deconvolution). MTF must drop to 10% before detail truly becomes extinct, and for a 2.2 micron pixel like this 120 MP Canon, f/5.6 will still give you 18% MTF, and there are a host of lenses that are very sharp at f/5.6.
For further consideration, look at the effect of the anti-alias filter, which drops MTF of spatial frequences far lower than needed to suppress aliasing. The ideal solution to this problem is pixels that are so small that diffraction itself anti-aliases. That will increase contrast at lower spatial frequencies by 30%.
Instead of worrying about "wasted pixels", we should be worrying about "wasted glass". A simple prime lens has about the same resolution and cost today as it did 20 years ago, whereas image sensor resolution and cost have advanced by many orders of magnitude in the same amount of time. Therefore we should be more concerned about that part of the system which is expensive and difficult to improve; not the part that is getting cheaper and better.
It's true that some lenses are already into diminishing returns, particularly fast or wide lenses. But many other lenses are not even oversampled with 2 micron pixel sizes, including many macro and sharp primes. Take a look at this example of a $400 macro lens using 1.2 micron (simulated) pixel size:
http://forums.dpreview.com/forums/read.asp?forum=1019&message=29826265
Dynamic range is the distance between clipping and noise. The standard engineering definition assumes a SNR of 1:1 as the lower bound, but few photographers can tolerate that much noise and usually prefer 4:1 or 8:1. Random noise sources add in quadrature so that downsampling the pixel count by a factor of four increases the SNR by a factor of 2. A better way of thinking about it is this: the raw data from an image sensor has a noise power that increases linearly with spatial frequency (level of detail). Higher frequency (smaller details) have higher noise powers. If you throw away the high frequency detail, the noise goes with it. In actual practice, there are many better ways to reduce noise than by throwing away detail, and in any case, many viewers will prefer a detailed but noisey image over a blurry but less noisy one.
Actually, the "sensitivity" (more specifically, e-/m^2) is generally the same across a huge variety of pixel sizes, thanks to microlenses. What is not usually constant is read noise (AKA "high ISO noise", sometimes also referred to as "sensitivity"), because although it does naturally shrink a little bit as the pixel size is reduced, it's not always in exact linear proportion with pixel diameter, and therefore the generalization that smaller pixels tend to have slightly more noise in low light.
I completely misunderstood that title. Here's what I thought it meant:
A company named "Capcom Implements" has accidentally "Lost" their entire database of "Beta Feedback" for their upcoming game, "Planet".
What if China hires Ace Ventura: Pet Detective to find the missing dolphins?
Assume the car approaching from behind will not stop. You probably could have avoided that accident.
Several years ago, I was idling at a red light in Idaho during the Christmas shopping season. Traffic was greater than usual, so the line at the stop light had 20+ cars, much more than normal. Screeching tires and a split-second later the back-end of my Camry was smushed to bits. The driver had just barely touched their brakes, but I thought that I could have avoided it if I had been paying more attention.
For a while I was careful to notice any vehicle that approached quickly. Shortly I realized that it was futile. Too many drivers race to their position and slam on the brakes at the last second.
It's impossible to detect danger when so many drivers incorporate it as a regular part of their driving behavior.