Why is auto exposure so useless?

[Petrus]

There is supposedly 30,000 images in the camera that is referenced when you press the shutter and the camera chooses one of them. What I don't understand is if you have a setting of +1 EV or a setting of -1 EV are there any images that correspond to any of the two settings?

Slight misunderstanding here: there are no 30000 images in the camera to choose from, the multi matrix exposure algorithms were developed by analyzing 30000 sample photographs, then coded on the exposure meter.

If you have exposure compensation set, the camera simply calculates the exposure as usual, then applies the correction as set.

[stamper]

Slight misunderstanding here: there are no 30000 images in the camera to choose from, the multi matrix exposure algorithms were developed by analyzing 30000 sample photographs, then coded on the exposure meter.

If you have exposure compensation set, the camera simply calculates the exposure as usual, then applies the correction as set.

The Nikon site indicates that there is 30,000 images on board?

Quote

comparing what is sees against an onboard database of over 30,000 images

unquote

https://support.nikonusa.com/app/answers/detail/a_id/13774/~/what-is-the-difference-between-spot,-center-weighted-and-matrix-metering%3F

[AlterEgo]

There is supposedly 30,000 images in the camera

may be 30,000 datasets from matrix metering sensor, not images from the regular sensor itself.

[razrblck]

From the same link you posted:

What is the database of over 30,000 images?
Over the years Nikon has studied the color, area of coverage, focus distance, contrast, size and shape of shadows and highlights and exposure characteristics of over 30,000 actual photographic images and incorporated this data as a reference source for the expert exposure system that is the 3D Color Matrix Meter.

It's not a database containing actual images, it contains data extrapolated from those images that can be compared with that the metering sensor sees to find the closest best settings.

[stamper]

comparing what is sees against an onboard database of over 30,000 images for unsurpassed exposure accuracy,

unquote

It's not a database containing actual images, it contains data extrapolated from those images that can be compared with that the metering sensor sees to find the closest best settings.

The statement on the Nikon site seems pretty unambiguous to me?

[razrblck]

[stamper]

It looks as if we will have to agree to disagree about this? The bottom line is that the camera consults a database of information that personally I don't trust and prefer center weighted meter

[dwswager]

It looks as if we will have to agree to disagree about this? The bottom line is that the camera consults a database of information that personally I don't trust and prefer center weighted meter

Not sure what camera(s) you use, but my first question would be do you calibrate each camera when you first get it?  That is, do you use a meter you trust or at least an "exposure standard" that you want and then calibrate to either of those.  Most of the prior discussions assume the meter is actually accurate to begin with.  It has gotten better over the years, but I've had new cameras 1/3 to 1/2 stop off out of the box before.

Everyone has preferences of what exposure is "correct" for a given circumstance.  I use all 3 metering modes in my Nikons at different times depending on conditions and also use a Sekonic incident meter.  Even then, an accurate exposure may not be the one I want.

Bottom line:  Because of the subjectivity and sensor response differences, only when and if one could program in their own preferences would "auto" be more likely to match your expectations.

[Zorki5]

may be 30,000 datasets from matrix metering sensor, not images from the regular sensor itself.

This interpretation is IMHO closest to what's really going on.

Nikon's wording does not leave too much room for interpretation, and the way I read it is that the camera stores 30k of datasets for 1005 R, G, and B sensors (for the purpose of metering, only brightness is taken into account) along with distance information available for [some of] them (availability should depend on focus point coverage), so analyzing this whole database, in terms of processing power needs, should, in worst-case scenario, be similar to analyzing a single 30k * 1,005 == 30Mp image.

This should be doable with some concurrent processing in ASIC, and lots of shortcuts (ruling out some datasets early).

[MattBurt]

Saying that they "incorporated this data as a reference source for the expert exposure system" doesn't even mean there is a an onboard database, it means that the logic in the onboard system has algorithms based on data from those 30k photos which from an engineering standpoint makes a lot more sense than an actual onboard database of images.

I think a big part of modern photography is knowing the kind of conditions that make AE useful and what conditions make it unreliable. There are also a lot of degrees between those to extremes where AE will get you consistently over or underexposed and that is where compensation is handy.

[Zorki5]

Saying that they "incorporated this data as a reference source for the expert exposure system"

This is not what it says.

The link provided above (official Nikon support's info) has it as

comparing what is sees against an onboard database of over 30,000 images

[stamper]

Again this is the wording.

 an onboard database of over 30,000 images for unsurpassed exposure accuracy

A poor choice of wording???

[Slobodan Blagojevic]

It is worth noting that whatever it says, it's a translation from Japanese.

[Zorki5]

A poor choice of wording???

Or a marketing blurb.

Frankly... things like this matrix metering are usually done using simple neural networks. Sound complicated, but it isn't. The network would be trained with those 30k images, feeding data to those 1,005 inputs, and "telling" the network what correct exposure for each input set should be. This sets coefficients of network's internal nodes (computational layer(s)). Then, when you feed actual data from those 1,005 sensors, you get best guess at exposure. Implementation of such network itself is trivial.

[MattBurt]

This is not what it says.

The link provided above (official Nikon support's info) has it as

I was referring to the wording in that screenshot.
 
I'm a software engineer and while we may never know for sure if they could extrapolate that data from the images into a set of algorithms that may be the best approach from an engineering standpoint.  They might have a database of image data in there but I bet if they could figure out how to accomplish the task without it that would be more efficient (which in engineering often means better, especially on small devices).

Another thing I have learned in this business is the marketing department's statements regarding engineering aren't always accurate, but may be (someone's idea of) the best way to present a concept to end users.   

[SZRitter]

Frankly you don't seem to know how matrix metering is actually implemented for those cameras.

That's a little harsh, don't you think? He is just giving an example of how the data could be computed. That is all he is doing. Nothing he said contradicts anything else in other posts.

[Slobodan Blagojevic]

Someone off his meds today?

[Colorado David]

How many angels can dance on the head of a pin? The difference between including 30,000 images or including the data sets of 30,000 images is beginning to get a little tedious. Maybe I'm just cranky today. Shoot raw, expose the best you can, and do a good job of post.

[Zorki5]

Does Zorki5 seem to know how matrix metering is actually implemented for those cameras?

If you are asking about any level of certainty, then I'm afraid even Canon engineers responsible for similar systems in their cameras may not give you an answer that would satisfy you.

BTW, it's OK to ask any questions, but I'd appreciate it if next time you do it directly.

He is just speculating wildly. That is all he is doing.

That is 50% correct. I am indeed speculating/guessing, but I wouldn't say I'm doing it wildly. I am giving you an educated guess.

I have a US patent in image processing, I designed and implemented entire graphical subsystem for proprietary operating system (sold in early 2000th in the US as PDA FW), designed and implemented graphics formats, compression algorithms, libraries, you name it. Took Grand Prix of an international programming contest in 1993. Chances are you, Isaac, are using map or navigation software, graphics subsystem of which one of my teams helped to develop. Believe me, I can go on and on.

I find the subject of this thread very interesting, if not fascinating, and don't mind diving into tech details a bit. If this sounds like gibberish or out of place here, I apologize, but can't help. And I'd say that, with all my speculations, I am on topic in this thread, and you with your remarks are not.

If you think it's a waste of time, again, I can't blame you, but please note that this is exactly how I feel when someone  spends more than 5 minutes on the topic of subtleties of color management.

We could blether on about simulated annealing and genetic algorithms and fuzzy logic to as little effect.

You know what, that's an interesting proposition... 

[Zorki5]

Another thing I have learned in this business is the marketing department's statements regarding engineering aren't always accurate, but may be (someone's idea of) the best way to present a concept to end users.

Would agree.

That's what I meant when I said "marketing blurb".