It seems to me that DFD can't yet achieve its potential with current gen sensors. To use all feature of the lens that can move at 240fps, it is technically should be possible to have video that does not go focus hunting back and forth.
Jul 15, 2011 Jammin with the band. A big to everyone for their feedback on the original song/music video (Directed by Jay Ahn @j6ahn and Lo-Def @lodefbeats) and of course a HUGE thank you to the band: Breezy. Grabbing everyone's attention @ Crosby Beach. Antony Gormley's Another Place consists of 100 iron men standing on the beach.
However, there has to be enough light for 240fps. So say we are shooting at 60fps, then exposure should be bright enough that 1/240 is well exposed. So we shoot 1 frame from position 'a' at 1/240, then allow lens to explore forward for 'b' and 'c' additional shots and then immediately return back to make another 1/240 frame in 'a'. So each second we can have one primary focus sample and 2 additional focus samples - is it possible that our current M4/3 sensors just can't deliver enough read speed to get to 240 native fps? Or is it just CPU that can't process this much data?
Or may be sensor should allow sampling from any random rectangle within its area so that just that small area is analyzed for focus. We didn't reach pinnacle of DfD I think.
The thing is, there are a zillion options for subject tracking and prediction on the G9/GH-5 that were not explored in this article. Frankly, it is daunting to know what to use, so it could use some AI assistance.
But the default tracking is fairly conservative and slow. There are much faster rate options available, but I believe you then run the risk of going too far out of focus when the subject is lost.Also DFD performs very differently on different lenses.
For example it is very fast and accurate on the newer 100-300mm II lens I just purchased, and almost unusable on the 45-200mm I (read OLD) lens I have retired. It is also slow on the 20mm f1.7 ver I but works well enough. I assume on-sensor phase detect would be more consistent across all lenses in-system, and DFD requires lenses capable of supporting it, thus an annoying limitation. When commenting yesterday made an error, not for 'for each second' but 'for each frame'. So 2 measurements per frame or 120 measurements per second. In addition to primary full frame capture.
So should be about 180 focus measurements per second with lens that supports 240fps.However in insufficient light where 1/240 is impossible, that's where we get down to only 60 actual exposures and we see this hunting. I've analyzed bright daylight footage from my G9 and I don't see ANY hunting, but as soon as it is even somewhat dark - we get the hunting back. What I didn't test in G9 is if I force higher ISO to see if hunting would ago away at the expense of noise. I wonder if some future sensor can work in full readout 240fps at 4k then even frames that are used for focus hunting can be used for temporal noise reduction for each individual frame actually recorded, like HDR+ for video kind of approach.
Chris - great video, as always, but a few suggestions:I'd really appreciate it if dpr included screenshots depicting the various AF zones available on every model you review. Often these are not accurately depicted in instruction manuals, and video reviews seem to concentrate on eye-AF while ignoring other focus zones and modes.On another point, bright green squares dancing all over the EVF in response to subject movement can be very distracting, so it would also be nice if dpr could report whether these can be turned off, reduced in brightness or be assigned different colours.The viewfinder experience is at least as important as any other aspect of a camera, so it would be great if reviews presented more information and screenshots as guidance. DFD will always need more processing power than dual pixel. It needs more processing cycles to hone in. Dual pixel can provide (near) accurate depth information by measurement instead of calculation. Canon could do the exact same thing as dfd does if they wanted but with the aid of dual pixel info from the chip.
The only reason the s1 is quick for photo is an old trick, brute force powerrrrrr. Larger body, larger battery, more processing power. This also means the Lumix s camera's will always be a little bigger, more expensive etc. They have to include more power to do the same as pfaf and dual pixel. Did you watch the video? It does beat the dual-pixel autofocus in a few areas.It fails in video tracking.
Panasonic's focusing system continually beats other systems in speed and accuracy for Photography. The s1 beats the nikon z6 for tracking according to dpreviewtv.I think they should probably figure out something dramatically different for video, but their development has been about photography in the past, I'm sure if they put their engineers to solving the video auto-focusing problem they will make big improvements. It seems the arrival of good video auto-focus and great photo tracking is a relatively recent phenomenon. @cosinaphile nailed it technically.But the average smartphone use who tries a real camera will be disappointed from the user experience point of view.In a highly technological environment and at the doorstep to AI human tend to get more and more lazy and convenient. It's the way it happens.If I shoot a picture on my A6300 and on my iPhone 8 Plus of a landscape scene and share it on Instagram to Facebook the iPhone would win hands down with its right away edited picture.
The dull camera JPEG wouldn't stand a chance. What a great idea- a long overdue test imho. I can only say that I do industrial reportage work under very harsh conditions, and my G9 does a mighty fine job. I find the AF to be very reliable and super accurate. A few years back I found myself in the situation to produce some images during the Swiss Motocross & Sidecar Motocross Championships. That wasn‘t planned, I was there in my free time, and the only camera I had with me was my GX8, with my Pana 12-34 f2.8 and an Oly 45mm f1.8 lens.
Motocross is fast and dusty. The AF tracking worked so well and reliable that I got several hands full of great hits. With a camera that is anything but a sports camera. For still photography, I‘m more than happy with Panasonic‘s AF system. I find it to be very accurate, extremely fast, and it‘s my preferred AF system for everything but the most demanding sports shootings (which I don‘t do) or shootings in near darkness (which I don‘t do). M43 enjoy superior S-AF for photography, faster than Canon Nikon Fuji Sony in ten of 1sec snapshot. But M43 quickly fall apart when U need (continuous Video tracking) where it's C- AF was never reliable.
Today in 2019, the metric for AF has shifted toward Continuous Video. In this context, M43 is now at the bottom, even the once-horrible-Nikon now has a Sticky EyeAF that runs circle over M43. I owned & shoot M43, but it's is getting more uncompetitive everyday. It used be just the sensor, but now C-AF in video is also way behind the competition. The blame reside with Panasonic & Olympus for never addressing it's poor C-AF performance for years.
Poor C-AF was a problem when I had my GF2, it is still a problem in today's Gh5. @DamianFLI would hope that a 4.5k camera can’t be beat compared to the others which are about half the price or cheaper. But I would like to point out a situation where it can be beat.Very low light, when the sonys inevitably switch over to CDAF, and the EOS R keeps going full DPAF, and at very narrow apertures.
Shooting the A7III once at high noon, to capture a performance at the beach spontaneously, so I didn’t have any ND filters on me, resulted in very bad AF results. Shooting at ISO 50 and FPS set to 24, required me to have the lens stopped own to f16. With that DOF the wobble was extremely bad. Had to switch to manual AF. As we know sony’s AF system or rated to f11.Now that I have the R I was curious about its behavior.
Down to very low light. Stopped down to the narrowest aperture of f22. No issues.As someone mentioned though, for a subject that is jumping around like on a trampoline the Sony tracks better. For me low light is a more common situation. Great video chaps! When's Jordan moving into law enforcement?I have the Lumix G9. There are a variety of settings for use in tracking subjects available to me, across all focus types (single-point, area, custom, tracking, face-detect).
I've found that the default one works reasonably well at keeping a running toddler in focus with face-detect - the greater difficulty is actually keeping him in frame!I did have the opportunity to test tracking at the Royal International Air Tattoo in 2018, before the firmware update was out. Tracking worked, but I found AFF worked best in hitting focus on fast-moving jets.
If I get another opportunity, I'll try to work through some of the other tracking-setting available to me. Is it even possible to get pleasing results with AF when shooting video? Robotic focusing might not be pleasing to watch. It is like teaching computer to play music instruments.
There is so many things which will change in each situation. Try to code that. Even all humans don't agree what is pleasing.Panasonic should just focus to give more tools to aid manual focusing in video recording. Maybe some kind of AF scripts where you can preset different focus points, curves for focusing speed and change direction. It is like recording your focusing pattern and then execute it later.
This way you can carefully fine tune precise focus for different points. Might not work well with random action, but pleasing results often require some kind of staging. Sony has the best hybrid pdaf/cdaf system on the planet, but there are a number of shooting situations where e-mount does not use hybrid af at all, including with native e-mount lenses, it's strictly using ospdaf.you can get an idea of how well pure ospdaf works when using certain adapted lens situations on the a9; it can be stunningly accurate with af-c, due to the stacked sensor making up to 60 af/ae measurements a second.more measurements and less prediction is where future af systems are headed. It's tough when you have to negotiate a minefield of patents.
Canon's for DPAF, Sony (and others) for OSPDAF/Hybrid. DFD is a clever way of deriving the additional phase information quickly without direct phase detection using split-pixel or dedicated-row phase detectors. It has the virtue of not having phase detection row artifacting, like DPAF, but without the signal loss of splitting pixels.On the other hand, there's simply no substitute for a focusing scheme that can produce two dimensions of focus information instantaneously.It's not that Panasonic engineers aren't clever.it's more that, aside from using light field imaging (with all its image resolution problems), there's not a lot of alternatives to PD and CD based AF technologies. DFD is an extension of what Panasonic knows how to do quite well. I the DFD have suffered in a FZ1000 and now in an LX100, and honestly it is something that I do not want to suffer again if I can avoid it, it is not that it is useless in AF-C with its 'eternal search', it is that it gives me false positives with easy situations even for a simple contrast AF system. I made 15 photos to a tree located about 50 meters a few minutes before dawn, I got only one shot in focus I tried with spot focus and fine focus.
With another tree located about 150m in broad daylight something similar happened to me. It's really bad, I do not know how it will be in the flagships of the brand, but in FZ1000 and LX100 it stinks (LX100 even more). My focus point was on the treetop in both shots (one of them is on my Flickr). False positives has happened to me with LX100 (not with FZ1000), it is not a defect of my unit, I have tried several, it only happens in certain conditions and especially in the longer focal (70mm equiv), but it is very frustrating.
The search for the AFC is terrible, I thought it was a problem with my unit and I even contacted the Panasonic technical service, if you make a burst you will catch shots in the fuzzy moment of the 'breathing' of the focus. I have also noticed in both cameras that they are very obstinate when it comes to identifying a subject that is slightly darker than the background, however large it may be, in a detailed picture it costs me a lot (despite that '80s' button in the one that you choose AF Macro for close objects) the LX100 is even more stubborn than the FZ1000. But both do worse than a competitive model with 25 points of simple focus by contrast (despite its 49 DFD points).
DFD produces visible warbling during video recording because it requires two delta frames (plus a 3rd final focus-movement frame) at slightly different focus positions to perform its depth calculations and naturally those frames will be captured as part of the video. I wonder though if this could be minimized or eliminated by synchronizing the intervals of the DFD frames relative to video frames, in possible combination with over-oversampling the video frames so that the camera can toss-out OOF DFD frames. For example, when the camera is configured to shoot at 24p, shoot at 48p/60p instead (or 3x to eliminate both DFD frames), and based on the DFD calculation result, toss out the DFD frame (1st or 2nd) that is further OOF so that it’s not encoded in the resulting video. DfD can determine phase differences, depth map, and find focus with much fewer frames/iterations than traditional contrast detect, but it will still always need to take multiple frames to generate that information. This is from the original IEEE paper where DFD was first presented'A new method named DFDlF of determining depth(range) from image defocus and rapid autofocusing of acamera is presented. It requires only two images in theory(but three images in our implementation). 'Maybe Panasonic is still using more than two images for a given focus calculation, so there could be room for significant improvement, but it's never going to be like phase detect where you get your phase information from a single frame.
So the outcome of the video is that the GH5 with the faster processor keeps up with a SL3?Please don't sugar coat the issues around DFD and the Panasonic cams. There is a reason why those with Panasonic cameras shoot video on MF instead of relying on AF as with Sony, Canon, Fuji and EVEN Nikon.Panasonic can try as they want to say DFD is great but its just GOD AWFUL as it was plenty evident on the video. I know you said you had only those 2 cams on hand but a better comparison is A7III, Z6, GH5, XT3 and EOS R.To be fair we all know the outcome here of such a test.
Canon/Sony leading the way with Nikon/Fuji slapping each other for third place and the Panasonic asking when is the race STARTING.Panasonic AF is a PIG with lipstick. A pretty one, even sexy. One that I would introduce to my mom. Could have few piglets and perhaps grow old together. LOL!But in the end it is still a PIG.