Hacker News new | ask | show | jobs
by Sharlin 460 days ago
Nowadays they use pre-trained pattern recognition AI models, yes, which has become much more impressive (and CPU-intensive) with mirrorless cameras where the entire resolution of the main image sensor is available for analyzing the scene. Some higher-end traditional DSLRs have a "high"-resolution (around 0.1 MPix or so) metering sensor that is used to assist the AF system (eg. what Canon calls iSA and iTR [1]).

Traditionally, cameras would just focus using the single focus point the photographer has selected, or if they have selected a larger area focusing mode, the camera would typically pick the closest point of a group of points, assuming that that's usually what the photographer is interested in. (Remember that traditional (D)SLRs have a discrete AF sensor with at most a few dozen focusing points to choose from!)

In tracking AF modes (eg. Canon's Servo AF), depending on settings, the camera tries to avoid sudden shifts in focus even if a foreground object momentarily occludes the original target. Tracking AF also has to predict the subject's motion to prevent the focus from lagging behind a fast-moving subject. Higher-end cameras allow configuring the AF behavior in terms of how reactive vs "sticky" it should be when tracking a subject, and how linear the subject's motion is expected to be.

[1] https://www.canon.com.hk/cpx/en/technical/pa_Overview_of_65-...

3 comments

With mirrorless cameras the focus switched from specialized sensors to on-CMOS contiuous exposure sensors, so movement is easy to detect. At this point the cameras have specialized AI hardware to run the models, and they also accept user input (on R5 MkII you can register up to ten people to prioritize focus on[1]). The focusing options are now very complex[2][3], and combined with lots of customization options on the camera's buttons you can have very specialized/personalized setups for different types of photography.

[1] https://cam.start.canon/en/C017/manual/html/UG-04_AF-Drive_0... [2] https://cam.start.canon/en/C017/manual/html/UG-04_AF-Drive_0... [3] https://cam.start.canon/en/C017/manual/html/UG-04_AF-Drive_0...

Sure, as I said in the first paragraph, AF is these days very impressive thanks to the large amount of data available (but of course this would have been too much data back in the day, when there wasn't nearly enough CPU power to process it fast enough). I wanted to give more historical context for how AF worked before fancy AI.

The AF settings, except those related to face/object recognition, haven't actually changed that much since the 7D Mk II days. The preset system is more general now and allows you to store and recall all AF settings rather than just the three tracking-related variables. The high-end DSLRs used to have six cases for different types of sports that you could modify but not rename.

Sony A9 III even has a configuration setting of whether it should focus on the left or right eye of the person :) It also can remember faces and prioritize them if there other faces. Let's say someone shoots their kid on the football field and wants only them to be in focus.
The left eye/right eye option has been around even on lower end cameras for the better part of a decade, which is kinda wild. My fuji from 2017 has that, and Fuji are specifically known for having worse autofocus performance than Canon or Sony or Nikon. Nikon and Sony specifically seem to be top of the pile right now.

The Z8 has a whole separate processor dedicated to autofocus and the viewfinder which, in practice, means it can shoot 20 FPS full quality 45 megapixel RAW files with continuous 120fps autofocus without blacking out the viewfinder for each shot, which is absolutely insane.

The a9 iii shoots 120FPS 24.6MP raw files with continuous autofocus and without blacking the viewfinder. Top-end cameras are insane.
I sort of miss my Canon EOS 1 at 10 fps (film) when shooting sports. "If you saw it, you missed it."
The $$$ per second of shooting that is hilarious. I guess that’s what shooting IMAX feels like, but only for like 3 and a half seconds at a time.
The technical term for the "stickiness" you're referring to, where a system is resistant to change or has a memory, is hysteresis.