5

Reading answers to Why are my sky photographs coming out dark?, I began to wonder: are all current autoexposure algorithms really so basic as to just measure the average brightness of the scene?

It would seem that one could get much better results using for example some machine learning system that is trained with real, well exposed photos. It could then more accurately estimate the correct exposure from the preview image, even for the black cat in a coal mine and the white dog on the snow.

And I don't mean the specialized scene modes, as that's just moving part of the task to the user.

  • Modern DSLMs offer rather good exposure preview right in the electronic viewfinder, which tends to be good enough for you to notice if your camera is about to screw up the exposure.... – rackandboneman Jan 16 at 20:46
  • Apple's deep fusion on iPhone 11 does almost what you want but currently only works on medium to low light. It quickly takes multiple pictures with varying exposure then use software to combine the pictures to enhance details. I suspect you can tune such an algorithm to work with outdoor photography as well – slebetman Jan 17 at 5:11
  • An almost magical innovation with striking results available in any $100 pocket camera these days: face recognition. The camera identifies faces in the view and exposes them correctly. Because the people are almost always what we are interested in in an image (did you notice how uninteresting all the sunsets are 10 years later? What you are looking for in old pictures are invariably the people!) the results are almost always better. (But the skies are always white -- you need to use flash in order to get both). – Peter - Reinstate Monica Jan 17 at 15:16
15

are all current autoexposure algorithms really so basic as to just measure the average brightness of the scene?

No.

Do any cameras have advanced auto-exposure algorithms?

Yes.

It would seem that one could get much better results using for example some machine learning system that is trained with real, well exposed photos. It could then more accurately estimate the correct exposure from the preview image, even for the black cat in a coal mine and the white dog on the snow.

Most cameras have multiple metering modes:

  • Spot – The brightness of a small portion of the image is used to calculate exposure.

  • Average – The brightness of the entire scene is averaged and used to calculated exposure.

  • Center-weighted average – The brightness of the entire scene is averaged, but the center is given more weight.

  • Matrix / Evaluative / etc – This is the mode that attempts to do what you describe. The scene is divided into multiple parts. The brightness of each part is evaluated. The result is used to search a database or fed into an algorithm to determine exposure.

I'm not aware of any camera that currently does so, but matrix data could easily be fed into a neural network. That would allow a camera to learn its owner's exposure preferences. This is the type of tech Google uses to identify cats in YouTube videos.

| improve this answer | |
  • 4
    When using 3D color matrix metering (most) Nikons compare the metered scene to a database of over 30,000 image/scene types in the process of determining the proper exposure. – Steven Kersting Jan 16 at 18:41
  • 3
    @StevenKersting Pretty much all camera makers have at least some models with library based scene recognition. As one moves up through the models and their price points, they get increasingly complex and capable. – Michael C Jan 17 at 0:24
  • 2
    Note that all of this still uses the same meter which in most camera's is measuring reflective light so has still to be (manually) corrected for non grey subjects. And in really highly dynamic lighted scenes you'll be limited by the dynamic range of the sensor. Enough options for still getting bad exposures, even with smart meter modes. – hcpl Jan 17 at 9:55
4

Yes, lots of them do. But they still can't read the photographer's mind regarding which part of the scene is desired to be the most properly exposed.

Sometimes the camera can only be as smart as the photographer allows it to be.

Even back in the ancient film days, primitive multi-segment light meters allowed for very basic analysis of which parts of the scene were brighter and which parts were darker. When semiconductors were introduced into cameras' logic systems in the 1970s, this information could be used to do rudimentary scene recognition. If the upper third of the frame was much brighter than the lower two-thirds of the frame, the camera would bias exposure on the assumption the photographer wanted the darker areas in the middle of the exposure range. If the upper two-thirds of the frame was brighter and the lower one-third was darker, the camera would bias exposure based on the assumption the photographer wished to expose the brighter areas in the middle of the exposure range. This worked fairly well for landscape photography.

Over the years light meters in SLRs and then DSLRs have multiplied the number of discrete segments from single digit numbers to hundreds of them to thousands of them. They have gone from being truly monochromatic to dual layer (measuring and comparing brightness at two different wavelengths of light) to today's RGB-IR sensors that are effectively miniature color imaging sensors. With the advent of mirrorless cameras (and Live View in DSLRs), metering can be done using information from the main imaging sensor.

As data rates and memory capacity available to camera designers have increased exponentially, the complexity of "library based" exposure metering routines has also increased. As a result, cameras are getting better and better at recognizing many different types of scenes and adjusting recommended exposure based on that identification.

But cameras still can't read the photographer's mind, even if they are getting better at guessing what the photographer probably wants.

This is particularly the case when the photographer uses one metering mode, such as matrix/evaluative metering, in a situation where another metering mode, such as partial or spot would more accurately inform the camera exactly what part of the scene the photographer is most interested in exposing in the mid-tones between too bright and too dark. Or where the photographer limits blown highlights in any part of the scene by turning on Highlight Tone Priority (Canon)/Active D Lighting (Nikon)/Whatever other camera makers are calling it.

Sometimes the camera can only be as smart as the photographer allows it to be.

| improve this answer | |
3

Camera manufacturers and film processors have been working this issue literally since film was invented and people stopped processing it themselves. We've progressed from ambient light meters using solar cells to incredibly complex databases of exposures taken and analyzed by multi-point arrays in a flash 'pre-flash' pulse.

And yet... we still hose exposures.

Most exposure metering now adays I'd hazard to guess is centered around 'beautifying' the image. That is- compensating for exposure deficiencies and poor lighting conditions by actively modifying gain, saturation, and offset bias for various portions of the image. Those aren't so much metering as interpretive of the collected photons, but still-

There was a concept called 'paxelization' that I haven't heard outside of Kodak. It was akin to the modern day downsampled machine learning imagery. That paxel of pixels was small but was used by the various algorithms to predict the ideal exposure (for both film and digital).

So yes, cameras do do more than what you'd expect, and the software that renders your image or prints your film (digitally or optically) does as well. It truly is amazing.

I'm short on the answer here because, literally, books have been written about metering and printing. If you have interest in photography please consider reading Ansel Adam's books (2 of the 3) the Camera and the Negative. The Print book is useful too and will go into deep, deep detail how to apply what you've learned, but isn't so relevant to your interest in metering systems.

| improve this answer | |
0

Yes. In late 2019, Canon launched its flagship EOS-1D X Mark III DSLR for professional photographers. This camera flaunts the most exceptional self-adjust yet observed, together with very good quality video/film highlights numerous individuals expected would just be found in the most developed mirrorless cameras. Nikon has announced its own rival, the Nikon D6, which combines rugged, durable design with advanced autofocus and high-speed shooting.

| improve this answer | |
-1

Smartphones are computers with a camera sensor. Cameras are sensors with a computer. All this computing requires times and energy and would come with a reduction of battery life and shoot rate. It also requires a lot of storage which would be offline, so the camera would require a SIM card and network access...

And so far in the smartphone the computing happens after the shot(s) so is pretty much what you would achieve with exposure bracketing and post-processing

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.