Hi everyone, this is my first post, with my first drone, purchased over Black Friday (no significant discounts unfortunately), and I've been enjoying learning to fly it.
I opted for a Mavic Air 2 becuase I was attracted to the better camera on the Air 2 over the Mini 2. Having said that, I've been quite disappointed with the camera output since buying the drone, and I'm starting to wonder if I have a faulty unit.
I'm certainly no cinematographer, but I am a video engineer professionally, working with a focus on display and image accuracy. I do a lot of work with Hollywood, and am quite familiar with manual settings on cameras, and know to achieve a ~180 degree shutter angle I should be aiming to set the shutter speed to 1/50 for 24p, 1/60 for 30p and 1/120 for 60p captures.
I have tried capturing video in Normal and D-Cinelike as well as H.264 and H.265 and found serious compression artifacts with all combinations of settings.
I typically use DaVinci Resolve to edit and grade content. Here is an example of some content with slightly less compression artifacts that I captured just before Christmas:
I had to apply some pretty heavy noise reduction to achieve an OK result with that content, and this was captured in Normal mode H.265. It looks OK in a window on YouTube, but I'm not that impressed when viewing it at UHD resolution full screen on a monitor or TV.
What I'm noticing, is that when there is motion, the compression artifacts might be somewhat obscured. It's when the drone is still, than any surface with texture on it, pulsates at what I assume is the same rate as the GOP structure because it's about once / second. I'm also noticing coloured blocks of noise which look like they're YCbCr errors becuase they flicker between green and magenta, which is typical of a slight offset of Cb / Cr around neutral.
All of this noise is apparent in the raw out of camera footage, but is far more apparent once a grade is applied. It appears to me that D-Cinelike actually has more of this noise than Normal, but that isn't always the case.
Here is a good example of the colour noise that I'm talking about. The first image is the ungraded output from the Air 2, the second is the graded image with green and magenta blocks over the roofop region, and the third is a deliberately exaggerated version with saturation boosted only for that region. The issue is more apparent in video because the magenta and green blocks change each I-Frame.
The full original clip can be downloaded here: https://lgtca.box.com/s/2v22u2620zzjst251trozunw1fhiwk3p; it's 1.2 GB I'm afraid. I'll try to find a way to cut out a section without re-encoding to make that a bit easier.
I'm not simply trying to complain. I'm actually hoping that I'm doing something wrong, or that I might have a defective unit so that I can return it and try a different one. Overall, I had expected much better compression from 120 Mbps.
You will have to find the optimal settings for your workflow I guess.
There is a tradeoff in all choices I'm afraid.
I don't own a Mavic Air 2 so I can't tell you mine, but so far I have not seen a consumer DJI drone that delivers footage that is on par with my bigger camera systems.
The camera system (lens + sensor) on these small drones is just too small / lightweight to deliver great footage I guess.
Lots of (in camera) corrections that don't do the details any good.
But that is no different from my iPhone actually.
And that does not fly very well
Check youtube for 'best image settings Mavic Air 2' and you will find some hints to start with.
OK I did some more scientific testing yesterday. This was my process:
Hover the drone in a fixed location, capturing a detailed, but fairly low contrast scene; I used a quiet residential street on an overcast day
Lock White balance and use manual settings
Capture at 0, +1, +2, -1 and -2 EV
Capture the scene at 4K24 & 4K60
Capture in Normal and D-Cinelike
Capture the scene in with SRT subtitles enabled to identify exposure later
Load captured data in to DaVinci Resolve
Grade all content to match 'Normal 0 EV' as best as possible
The idea in doing this is to see if D-Cinelike provides any additional usable dynamic range, or reduced noise over Normal. I also hoped to see how far the image can be recovered from under or over exposure. Finally, I also wanted to see if there were any best practices for grading D-Cinelike that I could note down.
During this testing there were a number of interesting take aways, and I'll try to summarise these before providing more details.
D-Cinelike offers no significant advantage over Normal; no increase in Dynamic Range, and no greater ability to recover under or over exposure
D-Cinelike requires significantly more grading adjustments to achieve an acceptable result, particularly if exposure is too high / low
The DJI Fly software / drone / camera suggests different EV results for Normal and D-Cinelike with the same shutter speed, even though the sensor is capturing the same light
The suggested EV in the DJI Fly software / drone / camera does not follow physics; doubling exposure time does not result in 1 additional EV, and to achieve a +2 EV suggestion, exposure must be increased much more than twice
and finally
In practice, Normal mode offers lower noise, increased fine detail, and greater highlight and shadow recovery than D-Cinelike
D-Cinelike has colour block noise regardless of exposure
EV values cannot be trusted for D-Cinelike at all, or for positive values in Normal
Unfortunately, the day I filmed was quite ovecast, and I actually found the 'Normal' EV +1 to be closer to a well exposed image. Certainly, the sky was mostly blown out in this image, but the mid-tones and shadows seemed more accurate, and I therefore used this as the basis to grade all other EV / frame rate and captures to.
The first thing I noticed when consulting the subtitles, was that the EV values shown in the DJI Fly app are accurate when shifting down (less exposure; negative EV value), and grossly inaccurate when shifting up (more exposure; positive EV value). The shutter speed required to achieve an EV value of 0 is also different in Normal and D-Cinelike. Replicating the 0 EV value shutter speed from Normal in to D-Cinelike results in an EV value of +2/3, and this exposure seems to be a good starting point for D-Cinelike. The table below summarises the shutter speeds and EV values:
The next thing that became obvious, was that while it's possible to grade D-Cinelike to appear like Normal, grading the Normal mode content to match other EV values was really quite easy. Lift, Gamma, Gain and Contrast got me really close in matching the image.
After grading all of the clips to look similar, I went and looked for a the presence and severity of noise in each version. Obviously, -2 EV results in a very noisy image once lifted, the -1 EV Normal mode clips were also noisy, but might actually be usable with a bit of noise reduction (which I didn't do). Keep in mind that my reference actually became the +1 EV Normal clip, and not 0, so -1 needs a lift of 2 full stops to match! D-Cinelike noise levels are signficantly higher than Normal, and start to have these strange Magenta and Yellow blocks appearing even in non shadow areas. This portion of a D-Cinelike -2 EV frame is a great example of this:
Unfortunately this coloured block noise is visible in all EV captures in D-Cinelike, obviously reducing in visibility, but present even in the +2 EV frames, which is very frustrating and makes D-Cinelike effectively unusable.
+1 EV was my reference, and as I said, does have blown out sky, +2 EV, which as I show above is actually +3 2/3 EV, was totally unusable, as I would expect.
For more fine grain noise as opposed to the coloured blocks, D-Cinelike appears worse than Normal at similar Exposure Values. I believe the reason for this is that 0 EV in Normal is 2/3 of a stop more exposure that 0 EV in D-Cinelike... that's quite a lot less light that the sensor is having to work with. I believe that if the EV values in D-Cinelike matched those in Normal, the noise would be similar.
The usual approach of Exposing to the right, within reason, appears to work well this this sensor. Noise becomes more visible if you need to lift exposure at all, and this can be seen even in the Normal 0 EV when I grade to match +1 EV. Ultimately the scene I captured, even on this overcast day, is overwhelming the sensor. I can capture the full range without blown out sky with 0 EV, but the mid-tones and shadows will be too low, and raising them will amplify noise, or I can expose the mid=tones and shadows correctly and blow out the sky. I was able to apply a complex curve to the 0 EV grade, preserving the sky and increasing the mid tones and shadows, but it wasn't as easy as matching exposure, and did result in some noise in shadow regions.
I have exported two versions of all the 4K24 and 4K60 clips from Resolve, one with the grades to match, and the other without grading which just shows the raw camera data. Obviously these are re-encoded, but the bit-rate used is high. These files (420 MB) can be downloaded here:
In order to achieve good motion response, it is best to aim for 2x fps; that's what I meant by 180 degree shutter angle. I do usually aim to do this when actually capturing video for anything other than experimentation, yes.
In this case however, I was focussing on noise rather than motion, and the shutter speed is therefore only relevant in how it affects exposure.
Nezil Posted at 1-3 20:19
In order to achieve good motion response, it is best to aim for 2x fps; that's what I meant by 180 degree shutter angle. I do usually aim to do this when actually capturing video for anything other than experimentation, yes.
In this case however, I was focussing on noise rather than motion, and the shutter speed is therefore only relevant in how it affects exposure.
Actually a fast shutter can make the image look very noisy.
lvl.1
Captain
