Mini 3 pro not as good as expected?

[jphoto]

With regard to the claims that the Air 2S has a better low light camera....What am I missing? These videos, while not exactly scientifically controlled, and with no actual measurements, seem to make the Mini 3 look better than the Mavic 3:

These compare the Mini 3 to the Air 2S, and once again, the Mini 3 seems to look better:

For a specs comparison, this:

DJI Mini 3 Pro vs Air 2S [Insider comparisons]

In this article I will share with you the differences between DJI Mini 3 Pro vs Air 2S So that you can make the best decision for your next purchase.

droneflyingpro.com

Though the good doctor seems to think the mini 3p sensor is larger than the one in the Air 2S. Investigating that one now.

 

[jphoto]

So here was my question: What is the actual size of the image sensor in the Mini 3 Pro? The Air 2S and Mavic 3 is directly specified (1" and 4/3 respectively). The M3P sensor size is specified as 1/1.3. What does that mean? Well, roughtly, it's an odd way to specify the diagonal dimension of a sensor relative to 1". So 4/3 (in the Mavic 3) is 4 3rds of an inch, or 1.33". That means the M3P sensor is 0.77" or 51/64 of an inch (you see why they wrote it as 1/1.3). But that's only the physical size, the actual pixel pitch is not specified for the M3P, and frankly, isn't really specified for the other drone cameras either, with DJI using the term "effective pixels". And this is where it gets confusing. "Effective Pixels" directly implies they are not "actual pixels", but the result of image processing. With the M3P sensor physically smaller than the A2S, but having more "effective pixels" by more than double, that should mean the pixels are smaller, capturing fewer photons, and producing a noisier low light image.

But that's not what we see. So, there has to be some tricky image processing going on, otherwise the A2S should win this hands down. Remember, the lens in the M3P is only 1 1/3 stop faster, but the sensor is smaller, AND there are more pixels, so it should be much worse. It's not, so we aren't being given all the information needed to understand this. For now, the proof has to be in the casual tests and video samples we have.

This graphic kind of sums up the advantages to each, though in some cases the differences may be quite small, even negligible, for an individual's purposes (source is droneflyingpro.com):

 

[akdrone]

Uh, yeah. A fixed aperture of f/22 will have a hard time capturing ANY light...

but hey...the darkness will be in focus!

 

[jphoto]

Forget f/22, just join Group f/64 and be done with it.

And remember, when you get to high f stops it's only a matter of time.

 

[ziv2liv]

My big issue with the M3P is auto-focus, and I think it's not the camera but rather the gimbal. I shot several frames at 48mp that look fine until you enlarge the photo (Zoom in), than you notice fuzziness. The reason I suspect it's the gimbal is because when the drone is grounded, the photo seems fine, but when in the air, especially at higher altitude, you'd notice it. I think that the gimbal cannot keep up with the drone movement fluctuation even when it's stationary in the air . The drone perhaps is too light.

 

[jphoto]

My big issue with the M3P is auto-focus, and I think it's not the camera but rather the gimbal. I shot several frames at 48mp that look fine until you enlarge the photo (Zoom in), than you notice fuzziness. The reason I suspect it's the gimbal is because when the drone is grounded, the photo seems fine, but when in the air, especially at higher altitude, you'd notice it. I think that the gimbal cannot keep up with the drone movement fluctuation even when it's stationary in the air . The drone perhaps is too light.

Interesting! Can you post a sample or two (full size of course)?

 

[rickmurray1989]

I fully admit that I approach drone photography/videography than a lot of folks in this forum. Although I am FAA 107 certified and experienced, I am a photographer first, and I capture images and video from that perspective. Sensor size is important, however, there are a lot of tools and techniques that you can utilize to get the most out of your drone images and videos. In the case of stills, you would be well served by capturing your images in DNG (RAW) format and processing them in a good RAW processor, e.g. Lightroom.

 

[jphoto]

It is a strong suspicion that the 48MP is not "real" but rather the result of some form of processing done internally to the real 20MP image data. On a basic level that implies interpolation. Yes to RAW capture, in any and all cameras. Lightroom is good, Photoshop is the same, but I recently discovered Topaz. Wow. It's what Adobe should have done. AI processing can be incredible, and it can be fooled too. So far, Topaz Gigapixel, Denoise and Sharpen beat everything in Photoshop so long as you're working with real digital images. Scans of chemical media...not so much. Fortunately, no chemical media on drones, at least since the Camroc (googling may now commence).

 

[tangogal]

The engineering and design that went into the Mini 3 Pro is the best I have seen from DJI in years.
To much emphasis is placed on 249g weight. Ridiculous that weight is a recipe for flyaways if you are not extremly careful. I installed the higher capacity battery in mine which added 70 grams however, now the drone is far more stable in the wind. It flies so long I need a chair.

I also added the higher capacity batteries so I registered the M3P with the FAA. Big deal. The cost was $5 and I printed a label with their registration number to put on the drone. I am a new flyer and have not tested the drone at night as yet. What I do wish is that DJI had put a fan in the M3P as it does tend to overheat when hovering. So far, that is the only negative I have found in the few times I have had occasion to use it. A fan would have added to the weight but so what? Registration was on-line and a breeze.

 

[Droning on and on...]

It is a strong suspicion that the 48MP is not "real" but rather the result of some form of processing done internally to the real 20MP image data.

There are 48M real photosites, but the sensor does not capture an RGB triplet of values. Rather, each element captures red, green, or blue, and the other two values for that pixel are interpolated from neighbors via a mathematical process called demosaicing.

 

[Eagle Eye 62]

Hopefully a lot of these responses aren’t guess work and based upon real world knowledge . I own both the Air2s and the mini 3 pro and the mini 3 pro has a better aperture for low light than the air2s and performs accordingly - The Air 2s has a slightly bigger sensor but higher aperture therefore not allowing as much light to the sensor - So in short don’t buy an air2s for better low light performance you’ll be disappointed if you’re not happy with the mini 3 pro - Consider reviewing your Settings or replacing your unit - The mini 3 pro performs well in low light

That has been my experience as well. Nit to mention there are plenty of Youtube reviews that herald the low light performance of the M3P. It sounds to me like you have a defective unit.

 

[jphoto]

There are 48M real photosites, but the sensor does not capture an RGB triplet of values. Rather, each element captures red, green, or blue, and the other two values for that pixel are interpolated from neighbors via a mathematical process called demosaicing.

We actually don't know that. The specifications say "Effective Pixels: 48MP", the actual number of photosites whould have to be much larger.

What follows is just my own technical musings with the goal to improving understand for myself and others.

So, some definitions are in order.

Photosite: a single light sensing element that responds only to light intensity (not color). To represent color it takes 3 photosites, each with a color filter. 3 photosites are combined using a demosaicing algorithm that combines the 3 sensors into one color pixel.

Pixel: The smallest element of a raster (grid) image. A pixel can contain color data or just intensity. The images we deal with from digital cameras have pixels with color data.

Effective Pixels: This is a means of stating how many pixels (color) actively contribute to the final image. There are some pixels that are outside the active image area that do not contribute to the image directly but may be used in calibration. These are not typically counted as "effective" pixels.

Total Pixels: A statement of the total number of pixels available on a sensor, usually a larger number than the "effective" pixels, though often the two numbers are very close.

Image sensor size: Stated in inch-related measurements, and refers to a 1" video camera tube's physical diameter. So a 1" sensor is the equivalent of the image sensing area of a 1" diameter video camera tube, but there is no dimension of that sensor that is actually 1 inch! The size of a "1 inch" sensor is actually more like 16mm diagonal. The chart below comes from Commonlands.com.

The number of pixels in a given sensor size directly relates to how many photons a photosite can capture. There's no real magic here. A lot of photosites in a small area capture fewer photons individually, and the result is more noise in low light conditions, and fewer photosites in the same area, are larger and capture more photons, with the result being a higher signal to noise ratio.

Based on physics alone, the camera in the Mini 3 Pro should produce worse low light images than the camera in an Air 2S. But it doesn't, so something else is going on.

One clue comes from closely examining (enlarging) a DNG image file. Weather has turned against me, so I couldn't shoot my own, and these are found online.

So...two images, the left from a Mini 3 Pro 48MP DWG file, the right from a Canon 5DIV raw file.


Original image files are here:
5D IV image
M3P image

These sensors are nothing alike. The 5D IV has a 30.4 Megapixel full-frame CMOS sensor measuring about 43mm diagonal, the M3P has "effective pixels" at 48MP (more than the 5d), but with a sensor size somewhere around 9mm diagonal. That means the photosites and resulting color pixels are much larger in the 5D, but the total number of pixels is significantly lower than the M3P.

But that's not just what's going on. Look closely at the results of image processing.

Both samples are enlarged 800%. There are clearly more pixels in the M3P image, but processing has blobbed them together obliterating the fine detail. No matter how much blow-up is used, the 5D image always has better detail, and fewer pixels.

It's just a theory at this point, but it may be possible that some form of smart pixel clustering is being used to greatly improve the signal to noise ratio in the M3P image at the expense of true 48MP detail. This would make sense. There was a need for a smaller camera for the M3P, but DJI would certainly want to improve its performance in every way. That means more, but smaller pixels. Smaller pixels require more noise reduction. There's probably a sweet spot where noise can still be effectively processed out, but resolution is still high enough to look good.

I hope nobody gets the wrong idea here, I'm not really trying to compare a tiny drone camera to a full frame DSLR, that's not the point. The goal is to try to decifer what DJI is doing to make a smaller sensor with even smaller photosites appear to work better than it should.

Philosophically, it's always about the right tool for the job. Drone cameras are no exception. We'd all like to see a little sub 250g aircraft have all the features we could ever want, but that's not practical. It does quite well for its size, and beats larger aircraft in some ways, not in others, beats smaller aircraft in many ways, not in all. It's a work in progress, so our complaining is partly justified, but also the product of impatience. I like it, I fly it a lot....and I like my Air 2S too...and I'd like to own the top-end Mavic 3 too. But I also know that as good as that might be, for photograph and video, my tastes are always toward the pro stuff, which is completely impractical for me now. So, Mini 3P it is.

As time and weather permit I plan to shoot my own test shots. I've literally been rained out for a couple of days, and now I'm back to work.

 

[RogerDH]

We actually don't know that. The specifications say "Effective Pixels: 48MP", the actual number of photosites whould have to be much larger.
...
Photosite: a single light sensing element that responds only to light intensity (not color). To represent color it takes 3 photosites, each with a color filter. 3 photosites are combined using a demosaicing algorithm that combines the 3 sensors into one color pixel.
...

The sensor has 48 million photodiodes which produce 48 million pixels. The demosiacing algorithm doesn't combine 3 photodiodes into one pixel; it guesses at the two missing color values for each pixel by looking at neighboring pixels.

 

[jphoto]

The sensor has 48 million photodiodes which produce 48 million pixels. The demosiacing algorithm doesn't combine 3 photodiodes into one pixel; it guesses at the two missing color values for each pixel by looking at neighboring pixels.

Yes, of course, sorry I kind of blew the explanation. The demosaicing algorithm interpolates (not combines) the data betweend different adjacent color photosites to produce full color pixels. The problem is, we can't just say that the image results in the equivalent of 48 million descrete color pixels, because different algorithms accomplish the goal differently, hopefully the best for the application. The pixels might be there, but they could be clumped enough to hamper their effectiveness.

It may be a moot point, though, since RAW files don't have the demosaic algorithm applied yet, at least that's my understanding of RAW in principle. And if that's true, what algorith is used actually could be a user choice (Raw Therapy, for instance). But it takes demosaicing out of the discussion of how to get low noise performance out of tiny photosites.

The fact remains that the photosites in the M3P are tiny as compared to the A2S, which should not result in better low light performance. Since the opposite is true, there's something pretty significant that we don't know.

If anyone has specific information on the exact image sensors used, that would be helpful here.

 

[RogerDH]

...
The fact remains that the photosites in the M3P are tiny as compared to the A2S, which should not result in better low light performance. Since the opposite is true, there's something pretty significant that we don't know.

The Mini 3 gets good low light performance from its f/1.7 aperture lens, compared to f/2.8 for the Air 2s. In 12MP mode, the 4 pixels that are "binned" to make one pixel have a larger area than the Air 2s pixels, and the process also reduces apparent noise. (The noise is still there, but the binning tends to average it out, making adjacent pixels more alike, so the noise is less apparent.)

 

[scro]

Going back to the "comparison" in post #32 it would be good to remember that a full frame DSLR typically has a high quality lens and "stresses" the optics much less than a small sensor camera like on the M3P. When using a tiny sensor you need an extremely sharp lens since the spacing between photosites is so small. On a large sensor you can get away with a much less sharp lens and not notice it, since the area the image is projected onto is so much larger. On top of that the very wide aperture the M3P has asks even more of the glass. What aperture was used for the sample picture from the Canon 5D (original picture won't open for me to be able to check)? I suspect it was stopped down a good bit lower than f1.7.

 

[jphoto]

The Mini 3 gets good low light performance from its f/1.7 aperture lens, compared to f/2.8 for the Air 2s.

Yes, it has a 1.3 f-stop advantage.

In 12MP mode, the 4 pixels that are "binned" to make one pixel have a larger area than the Air 2s pixels, and the process also reduces apparent noise. (The noise is still there, but the binning tends to average it out, making adjacent pixels more alike, so the noise is less apparent.)

The video comparisons are done at "4K", 3840 x 2160, not 12MP, though that would be quite interesting to see as well.

So...how large are the pixels in each camera? We don't have a lot of data on the Mini 3 sensor, but since it's specified at 1/1.3" relative to a 1" camera tube, and if we assume the actual image size follows the same ratio of tube-referenced dimension to actual dimension, we can calculate the horizontal width at 16:9 to be 10.73mm using data from the chart.

Pixel size is roughly calculated by taking the horizontal dimension divided by the horizontal pixel count (effective) times 1000, not making any allowance for other factors that may influence pixel pitch.

The Air 2S has a real image width of 13.95mm, divided by the pixel count of 5472 and scaled, we get 2.54um. Spec is 2.4um, so we're close to right.

The Mini 3 has a real image width of 10.73, divided by the pixel count of 8064 and scaled, we get a pixel dimension of 1.33um. If we bin 4 pixels in a 4K/3840x2160 image, we get a pixel size of 2.66um (2x horizontal and 2x vertical), slightly larger than the real Air 2S pixel size.

When you add two non-corellated noise sources the noise voltage increases by 3dB. When you add 2 corellated and equal signals, the signal voltage increases by 6dB. When this is done with 4 signals and noise sources, the total noise gain is 6dB while the signal gain is 12dB, improving the signal to noise ratio by 6dB over a single source, similar to doubling the light intensity. To that add the 1.3 f-stop advantage, increasing the light to the sensor by a ratio of 1:2.7, in other words, a little less than 3x the light. So there's the performance increase, but also, it's confined to a specific final image size. It's important to compare apples to apples.

We don't know how the actual binning in the Mini 3 is done either, because while 12MP is a nice, even 1/4 of 48MP, the actual image being shot is more like 8MP. It could just be a 4x bin and image crop, or something else might be going on. We could figure this out with a few test shots.

I haven't tried it, but I'll guess that a 48MP RAW still from the Mini 3 will look worse than a 20MP RAW still from the Air 2S in low light because it no longer bins pixels, and the Air 2S real pixel size is significantly larger.

 

[RogerDH]

...
The video comparisons are done at "4K", 3840 x 2160, not 12MP, though that would be quite interesting to see as well.
...
We don't know how the actual binning in the Mini 3 is done either, because while 12MP is a nice, even 1/4 of 48MP, the actual image being shot is more like 8MP. It could just be a 4x bin and image crop, or something else might be going on. We could figure this out with a few test shots.
...

The Mini 3's 4K video is a 3840 x 2160 crop of the 4000 x 3000 binned image. (The reduction to 8MP is mainly because video has a 16:9 aspect ratio, while the sensor and the photos are 4:3.)

 

[jphoto]

The Mini 3's 4K video is a 3840 x 2160 crop of the 4000 x 3000 binned image. (The reduction to 8MP is mainly because video has a 16:9 aspect ratio, while the sensor and the photos are 4:3.)

Seems to be a reasonable assumption.

 

[Droning on and on...]

Are you all aware of the dual-iso feature of the sensor that can be used in 12MP video mode that are labeled HD (re: HDR) in the settings?

This feature is responsible for the low light performance of the smaller sensor in the M3P vs the A2S. Capturing a 12MP frame, every other row of the 48MP sensor is energized for alternating high/low iso, then combined to produce a high dynamic range 12MP image.

Shoot some video without HD on the M3P (60fps I think) and you'll see better low light performance on the A2S.