Testing the Resolution of the Human Eye | Hacker News Reader

Testing the Resolution of the Human Eye

ipvm.com

29 pointsbrk44 comments

This is misleading on several levels. The size of an image (measured in MP) has almost nothing to do with the angular resolving power of a camera, and even less to do with the resolution of the human eye.

The article didn't mention that every camera tested likely has a slightly different focal length / field of view. Two cameras with the same resolution but different focal lengths will score very differently in these tests. For example, a 32mm lens will see much more detail on the chart than a 14mm lens, even with the same resolution sensor.

Additionally, none of these cameras have the peripheral vision capabilities of the human eye. Our brains pick up quite a lot of information outside the tiny rectangle that these cameras are able to capture. The resolving power of the eye also changes from the center of our vision to the periphery - just because we can read the E on the chart in the middle of our eye doesn't mean we can read it from the edge of our vision.

Comparing eyes to cameras at this level of simplicity is mostly meaningless.

They mention that the camera has about 60-degree horizontal FOV. A 5MP 16:9 camera is about 2980x1680 pixels, so about 50 pixels per degree.

Agreed, a eye to a camera is not a comparison, I've worked for a bit in computational vision after my PhD.

Our foveal vision makes the connection to digital camera sensors very unsteady. What we have is a spatial filter with diminished acuity as we move away from the focus of our vision.

We're actually processing what we believe to be a small amount of information in visual cortex. By the time objects make their way up the ventral visual system, they are encoded by very specific neurons firing for just that object. The brain is also thought to fill in a huge amount of information of our visual field.

You can't really leave the brain out of the equation when you estimate the sensitivity of our retina.

Also seems to completely neglect that digital camera sensors perform wildly differently in low light situations independent of the actual image resolution.

He doesn't mention it but it is fairly obvious in the last test where in low light the best cameras were "blind".

People don't often realize how poor our visual acuity is everywhere outside the fovea, the area of sharp central vision a few degrees wide (that's a few fingers held together at an arm's length). We need saccades and microsaccades and lots of processing power and plain guesswork on part of the brain to patch together the image we "see". The resolution of a camera sensor, on the other hand, is not a function of the angular distance from the center of the image.

http://en.wikipedia.org/wiki/Fovea_centralis

http://en.wikipedia.org/wiki/Saccade

http://en.wikipedia.org/wiki/Microsaccade

This is an awesome comment!

But can you see the entire line of that chart without moving your eyes at all? In terms of a single 'snapshot' of our vision, the resolution is atrociously bad. Only a very small area in the middle of our field of view has significant detail (test this by looking at some writing in the distance then stare at the first letter and try 'reading' the rest of the writing without moving your eyes - it'll just remain an indistinct blur).

Exactly. The fovea article on Wikipedia has an illustrative diagram: http://en.wikipedia.org/wiki/Fovea_centralis

If we're making it a competition, I'll note that the human eye natively supports edge detection, and it operates in a much wider range of light. Eyes also come pre-packaged in a format that allows for quite accurate determination of distance at no extra charge. While cameras have a slightly broader spectrum, eyes work in a far greater range of brightness - and seeing in the near dark is far more useful than seeing the LED on the end of a remote. Eyes are perfectly mobile, and don't require external power. Eyes have a much faster traverse and can move to and focus on a target much more quickly. If you're turning, eyes will give you a stable image of the world around you to help keep your reference - cameras will just give you motion blur. Eyes are very difficult to replace, admittedly, but they have pretty astonishing self-healing characteristics and last for decades of constant use - in the rain, snow, desert heat, even underwater (though admittedly the image isn't great). Not to mention that if you roll your cameras, people will just be confused as to your emotional state, and of course, it's harder to feel a knot of love when staring into your significant other's IP cam...

I've been reading the comments and something stuck out to me. It's not uncommon for people to have 20/15 vision with corrective lenses. The lenses don't zoom the image in, as far as I can tell; and so, the 20/20 number we often come up with has nothing to do with the resolution of the eye as a sensor, but more to do with the focusing power of the lens. It might be more useful to consider 20/15 or 20/10 as what humans can see, since the lens is whats failing on the eye, not the sensor in the back. To that end, too, it'd be nice to see some really good glass to mitigate the lenses failings there. There are charts that relate lens sharpness to sensor acuity. As long as the lens sharpness is approximately equal or better, we should have an accurate result.

Also, I have 20/15 vision with contacts in and both eyes open (weaker with just one eye open, as is normal); and, when I can see letters, I can see them pretty well defined. I'm not sure how much of that is mental post-processing, though.

it is very probably that using GIMP "curves" they would be able to bring a lot of details on 5MP and 10MP "blind" images. The digital cams lack dynamic range. A "blind" dark image is very possibly somthing like this rough simplified illustration - 0 where black and 1 [out of 255] where white. To us both values look black, ie. very close to 0 vs. 255. Using a tool like "curves" one can remap 1 to say 100 while 0 is still 0 and this will frequently make a reasonable, though somewhat noisy image.

Or in other words, they judged performance of camera vs. eyes by using human eyes. It would be interesting if they send the images to, for example, OCR.

Very interesting. However this only tells half the story about human vision, what about color depth, dynamic vision range, and speed (flash something fast on screen)?

Agreed. In this test, we were just looking at making out details.

I would certainly not say a 10MP IP camera is equal overall to the human eye. Indeed, to the contrary, the poor low light performance of the 10MP compared to the human eye is just one example of the tradeoffs of cameras to eyes.

The concepts don't map directly. The retina has a "static" contrast ratio of about 100:1 or 6.5 f-stops. However, the iris can change the amount of light that enters the eye, and the retina can change its sensitivity, so in practice it's more like 14 f-stops quickly and 20 f-stops if you let your eyes adjust for a while. That's not even counting the fact that your fovea has lots of cone cells while your peripheral vision has more rods, which is why looking next to a star makes it look brighter than looking right at it.

Persistence of vision means that flashing something on the screen is almost meaningless. All that matters is how much light energy is pulsed into the eye, creating an afterimage.

Half? More like 1% or less. Dynamic range is incredible (1,000,000:1 according to http://en.wikipedia.org/wiki/Human_eye#Dynamic_range), one should test vision in darkness peripherally (rods have lower resolution, but higher sensitivity than the cones in the fovea. There are claims humans can see individual photons), etc.

Even limiting one to these experiments, test subjects should dark adapt for at least one hour before a measurement (given that te article doesn't mention it, I think it is unlikely that was done). That can make a huge difference.

Some (and more, including "The brain contains 10^12 neurons, 10^14 of which are in the cerebellum.") of the numbers you ask for are in http://white.stanford.edu/~brian/numbers/node2.html

The 1,000,000:1 (20 EV) dynamic range is not that incredible when taking into account that it's dynamic, not static - it's the sensitivity difference between fully dark-adapted and direct sunlight, not the ratio of maximum and minimum luminance that can be simultaneously perceived.

My DSLR can do more than 20 EV of dynamic contrast ratio quite easily: first, it has a static contrast ratio of at least 8 EV (a conservative estimate); most lenses have about 7 stop difference between their minimum and maximum aperture; and my camera's ISO sensitivity range is 100 to 12800, that's another 7 stops. Together, these add up to 22 EV (a ratio of 4,200,000 to 1!) A high-end camera could probably easily reach 26 or 28 EV. And cameras don't need to a hour-long adaptation period.

Is the neuron number meant to be a joke, or is the 10^14 being greater than 10^12 a typo?

It is a joke, and an indicator of how little we really know. In the same vein, that text has:

Number of neurons in the brain in 1974: 10 billion

Number of neurons in the brain in 1994: 100 billion

(Wikipedia gives 85 billion on http://en.wikipedia.org/wiki/List_of_animals_by_number_of_ne...)

As the editor of the post, I'd love to get technical feedback and happy to answer any questions.

Why is there a bright outline around the letters? It looks like the photos have been sharpened.

I am not sure what you mean by 'bright outline'. We added red rectangles around the lowest discernible line. Are you referring to that?

We have not sharpened or enhanced the images. They are straight from the cameras, we simply digitally zoomed and cropped the portion of the eye chart to make it easier for people to see the results.

For example, in the "1080p - 160 lux" photo, at the bottom of the E there is a brighter line.

Edit: according to my color picker, the color of the chart is generally 8e9fb9 but goes up to becae4 at the bottom of the E.

If was shot in JPEG, the camera might have sharpened it a bit before even saving the image.

Were the cameras saving raw images? Looks like JPEG artifacts are visible in some if the images.

More information about the cameras would be nice. How do their lens sizes compare to the human eye? What kind of post-processing happens in the camera (if any)? What are the characteristics of the CCDs used? Etc.

Did all of the cameras tested have the same size lens? A physically large lens should have better low light performance, regardless of the # of megapixels.

The whole experiment seemed strange. Some of the lines you said were unreadable, I could read. Also, some of the ones you said were readable, were unreadable to me. Seems pretty subjective.

I hate to be negative on interesting experiments, but this test is almost useless.

The noise of the sensor and the JPEG compression of the camera almost overwhelm the actual resolution in the test images. Additionally, the eye gets a larger amount of information over time during fixation, while the camera is only exposed once, with the resulting aliasing of pixels against the structure of the pattern.

1. As others sugges, if you have a RAW setting on your cameras, you should use that. If you don't, you should probably get one that does.

2. I'm going to disagree with the naysayers who say this is useless. You can partially defang them by pointing out that while this doesn't fully test a camera against a human eye, it is a reasonable test on its own merits and generally answers the question that people are actually asking when they ask "what's the resolution of the human eye?" (After all, the difference in capabilities goes both ways; sure, human eyes are a great deal more sensitive on many measures, but would someone please show me their eye's 10x optical zoom? (Hello to the 2040s!))

A further observation that "megapixels" is not useful would be called for, though. We're really comparing angular resolution for a certain set of settings. But we can accept "megapixels" as a slang term if we also fix certain other camera parameters to ones common to consumer cameras in the market.

3. Using eye charts is a good idea, I think. I've seen comparisons that try to compare rod density on the human eye, but that's really hard. I'd consider adding a note that standardized eye charts measure what humans can really see rather than try to compare two extremely different hardware platforms directly.

We are using IP cameras, that's our area of research / focus. As such, those types of cameras so not have a raw option.

We used H.264 with default quantization. We could have certainly used MJPEG, as these types of cameras support that but in our preliminary testing, it made no difference.

That said, I do agree that a completely uncompressed, raw, image could make a modest difference. For our domain, IP cameras, we just don't deal with raw images.

Among the many hats that I wear, one is optics designer.

A useful resolution test is to take pictures of alternating black and white bars, then use some sort of image analysis software to measure the contrast ratio, then draw a graph of contrast versus "frequency" which is inversely proportional to the spacing of the bars. You could define the resolution as the frequency or spacing where contrast drops to some arbitrary value such as 50 percent.

Contrast won't depend on light level unless the camera is doing some interesting post-processing.

It might be worthwhile to Google "USAF test chart" and look at what pops up under Images, because folks have conjured up test patterns that are useful for determining the contrast of an imaging system in one exposure.

What I think might be tricky is to come up with an objective definition of resolution or contrast that is amenable to both physical and physiological measurement.

[edited to add "inversely"]

I know the tests were done by a company selling IP cameras but there are many cameras available which would make better competitors to the human eye. For instance the Leaf Aptus-II 12 which produces images with a size of 80 Megapixels.

We do not sell any IP cameras. All we 'sell' is research and testing.

Our focus is on IP cameras only, ergo the type of cameras we selected.

Fun fact: The human eye sees in higher resolution for red and green than for blue. Blue light is scattered more as it passes through the eyeball (and as an evolutionary result, blue detectors are sparser on the retina).

There are also algorithms that convert the raw data from the sensor into a compressed stream. That would downgrade the quality of the camera.

Though, that does make me wonder if our meat software does something similar.

First sentence: All sorts of wild guesses and theoretical calculations exist about what the resolution of the human eye is, with 576MP a common claim.

576MP is not a resolution. Article immediately closed.

Resolutions are commonly measured in pixels. Or are you expecting something rectangular and ISO standard like 1920 x 1080?

Divide by pi, take the square root, multiply by two, we're talking about 27000 pixels in diameter.

That said, I don't agree with a lot of the methodology. If pixels are what we're going for, it only makes sense to me to count rods and cones.

No, I think the GP means angular resolution – i.e. pixels per unit of solid angle – which one might presume is the figure of merit here.

It's trivial to transform the units, and it's easier to understand megapixels or xxxxp. Terrible reason to close the article.

It's trivial to transform the units

…except they're not dimensionally compatible. You need to know the solid angle subtended by those pixels. That's like saying you can trivially transform 60 miles per hour into a distance. You can't, unless you have time.

More importantly, the solid angle in question is not obvious – are we talking just the fovea (area of the eye with highest density of cones [1]), the eye as a whole, or the solid area subtended by a standard computer monitor, or…?

[1] http://hyperphysics.phy-astr.gsu.edu/hbase/vision/rodcone.ht...

The angle is the field of view. Because of the eye's uneven resolution you need to have some idea of how they're handling the field of view no matter what.

It's like saying '60 miles per hour from Vermont to West Virginia'. There's some fuzziness about where exactly you put the end points, but you have a good idea.

Basically, using raw angular resolution would also be too vague without context, but if you have context then megapixels are sufficient.

1. Yes it is.

2. The 576MP number comes from taking the resolution of the very center of the human eye and then extrapolating to the entire field of vision. It's the panorama you could get if you had someone sit in a single spot for minutes on end, examining every direction. This number deserves no better than a bad mention, being used so much in misleading contexts.

That's a little more than 4k (8.3MP). I think I was expecting it to be a little higher than that.

Reply on news.ycombinator.com