07-13-2018 10:49 AM
Nonetheless, I don't see why software can't be written to search for a bird near the center of the field of view. I say this as a mathematician and computer programmer.
Sorry if this is a bit of a diversion from topic... but I am in tech as well, so your comment caught my eye. Google is doing quite a lot with its AI platform wrt photography. In this  case, they worked on using AI to sift through "street view" images and basically work out which clips would make "good" photographs. They've also been doing well in competitions against human photographers where people "vote" on images. And here  is an example where they are doing more with short video clips that are determined worthy of capture with their AI. I suspect that phones will be the faster adopters of AI technology for photography and professional-grade cameras will lag a bit due to the use cases being somewhat different.
The unfortunate thing about AI is that it has to be "trained" ... and the training requires loads and loads of data (which, fortunately, Google has in abundance.)
There are quite a few cloud-based products that can use AI / machine-learning to enhance product functionality in ways just not possible (or certainly not practical) for individual users or running on non-cloud computers simply because of the impracticality of it (individual users don't have nearly enough data).
Machine learning isn't "new" ... the algorithms being used are things developed back in the 1980s. But you're reading so much more about now because today, with the internet and cloud computing, it's possible to feed these systems enough data with processors that can churn through it fast enough to be practical. But there are "data lakes" and "server farms" involved ... not individual computers or individual storage devices.
When you use a voice-assistant (Siri, Alexa, Google, etc.), you'll notice these devices are all cloud-connected and that's because your phone or smart speaker doesn't have the data or capacity to learn... but the servers in the cloud do.
So the question is: Could your camera recognize the bird?
The reality of how this would have to work (and why it's not a viable strategy ... at least not yet) is that the camera would need to capture a sample image (or a few sample images ... for example 1 second worth of video) and send it up to a cloud server. AI could churn through it and recognize that there's a bird in the photo and also recognize where the bird is located ... directing the camera's focus attention to the correct spot.
But think about the problems...
(a) the camera needs to be not just wi-fi enabled, but internet-connected (currently you can have your phone connect to your camera, but the camera only has a private connection to the phone... it can't use the Internet)
(b) this network needs to be fast and reliable. It needs to work everywhere you could possibly imagine taking a photo (and you can imagine taking photos in a lot of places that are beyond the reach of wireless networks.)
(c) the servers need to be able to respond very quickly. Basically you want to half-press the shutter and, all in the time span of a few fractions of a second, have the camera record a short clip, upload it, get it analyzed, download the results, focus, and shoot... and it all needs to happen fast enough that the camera owner doesn't complain about the horrible lag time.
So there's two answers:
From a "could we have an algorithm that does this?" perspective, the answer is... sure can!
From a "technically can we implement an architecture that provides an experience that consumers would like?" perspective, the answer is... sorry, not yet.
07-13-2018 11:06 AM
"<<Select a fairly high ISO 800 to 1600.>>
ebiggs1.... Can you elaborate on this a little just for my own knowledge? Does that ISO range somehow improve AF sensor capabilities?"
No it doesn't.
I don't remember the context of my reply as I didn't dig back through the replies but this is to help keep SS higher or faster if you will. It has nothing to do with AF. Camera sensors only see light and dark. All shades of gray. Black is on one side and white is on the other side of the scale. This is the Raw data. Color is figured out and added by the camera electronics and software.
07-13-2018 12:24 PM
<<Select a fairly high ISO 800 to 1600.>>
ebiggs1.... Can you elaborate on this a little just for my own knowledge? Does that ISO range somehow improve AF sensor capabilities?
ISO is a "gain" (amplification of the data) that the camera applies only after the exposure is completed (shutter is closed). It does nothing to how the camera behaves prior to or during the shot. The only exception is that if you use live-view, Canon performs "exposure simulation during liveview" meaning if you dial the ISO up & down while in "live view" mode, you'll notice the image getting brighter or darker. But it's just a "simulation" being shown to you and doesn't accurately represent what the camera sees.
07-13-2018 01:26 PM
TCampbell: I don't think AI, massive sifting through data, or an internet connection is necessary. Software to analyze an image looking for circles and parallelograms has been around for quite a while. It's really not that complicated. I don't think it should be hard to add ellipses. I've worked on somewhat similar image analysis problems. I'm guessing that a fast chip and a few gigabytes of RAM would be enough.
07-13-2018 03:10 PM
"I don't think it should be hard to add ellipses."
This is all a camera sensor sees...........
It doesn't see shapes of any kind. Perhaps you can make a break through and improve how it works but this is all a camera sees right now.
07-13-2018 06:18 PM
You can develop a computer algorithm to look for shapes -- that aren't really AI or Machine Learning. I used to work as a developer of CAD/CAM software and I recall developing a function that tried to guess what part of a geometric constuct you probably *meant* to select based on where you *actually* clicked with your stylus. It may have seemed "smart" but in reality it was a fixed algorithm with no abilty to "learn". So it wasn't really AI.
For true AI & Machine Learning, the system does have to be able to "learn".
Anyway... here's a good video series on the topic. As you mentioned your into math, you may find an appreciation for what it is doing:
That's just the first video in a series ... each one is around 20 minutes. But after an hour or so you'll have a pretty good idea of what's going on.
07-14-2018 11:45 AM
An accurate AF system that can distinguish between a small bird and twigs is most likely not going to happen soon enough to help either of us!
07-14-2018 06:06 PM
I have a Canon EOS 5D Mark III, purchased in late 2016. I've taken at least 10000 photos with it. I use either the EF 100-400 IS II or the 70-300 image stabilizer lens. In either case, when I am photographing a bird that is perched on a branch and has background foliage, AF usually doesn't work. I have lost at least 70% of the oppotunities for such photos, as eventually the bird flies off. The %$^&#$ camera insists on focusing on the background or foreground. These are situations where I am reasonably close to the bird, so that it opccupies maybe 1/12 or 1/16th of the image.
This is beyond frustrating. I have experimented with many different AF settings and nothing seems to do any good.
Anyone else have this problem?
I have saved a custom shooting mode for shooting a bird sitting on a branch, surrounded by other branches. I have found using a monopod, or some other support, really helps.
I select just the center AF point. Never bothered with Spot AF. I use Single Shot focus mode. I also use Back Button Focus, so that the lens does not try to refocus when I press the shutter. The focus stays put until I refocus it again. It usually takes me 2-3 attempts to nail the focus where I want it..
07-14-2018 06:10 PM