Mavic Pro-2 vs Zoom and P4P - Line Skipping + Sensor Heat?

[Cliff_622]

Guys, as previously stated, Sony Exmor CAN bin on the sensor. It is available on many of their sensors and saves readout time or enables higher frame count by the cost of resolution.

...but that‘s really not the point here.

So you are saying that Sony EXMOR has programmable logic onboard? We do know for a fact that EXMOR does A/D column conversion right at the end of each row/column. You are saying it's also doing raw collection "processing" before it dumps that data to the image processor? It would have to process multiple rows and columns, calculate that and then dump it out.;...and I guess an image processor would be non the wiser?

Hmmm,....I dunno,...this doesnt look like something "Sony" did at the sensor collection level. It's looks more like a third party binning or line skipping deBayer.

It's very hard to say for sure.

The bigger question for DJI is; "....WHY?....."

CT

 

[chunky]

Previously stated by who? Vertical and horizontal line addition isn't binning- it has the same outcome with respect to grouping pixels but it is done differently.

Does not matter how YOU call it, Sony also calls it binning.

There is hardly any sense to read the whole sensor first and add, bin or integrate the data in the SoC later instead of improving SNR.
Today it’s all about getting higher framerates from the whole sensor or reduce the sensor readout amoumt.
That‘s one reason why technology like Exmor was invented.
Since this isn‘t the problem we can stop here now.
We have specified 4K resolution in FullFOV and only get something below 2.7K - that‘s the point.

 

[chunky]

Another question again - you both are still talking about Ambarella as SoC.
Maybe - not more.
As far as I know the teardowns only have shown a chip with a big H3 on it - doesn‘t look like what Ambarella did before for labeling.

Again, it doesn‘t matter why it only works as present, it matters that it doesn‘t work as specified!
That‘s the only reason why we could demand DJI to fix it, not because of binning or adding (simple (CMOS) switch between two green, red, blue sensels to combine them) or skipping or Ambarella or Socionext or own SoCs.
DJI can do what and how they want as long as they fulfill the specs that are valid for our purchase!
That‘s the point!
The point is to make much wind (forums, videos, ask DJIs service WHEN this will be fixed) and show we do not accept the current situation if we don‘t like it.
Nobody told us before that we wouldn‘t get Phantom resolution when we were offered 10bit on the other hand. Nobody told us that the Mavic is less professional.
We were only aware that we won‘t get 4K/60p since it‘s not in the specs.

 

[chunky]

One last example for binning on the chip. Different sensor but also Sony Exmor, so CMOS:

https://www.sony-semicon.co.jp/products_en/IS/sensor1/img/products/ProductBrief_IMX350_20171109.pdf

„pixel BINNING readout funtion“ - don‘t tell me it doesn’t happen on the chip ;-)!!

Over and out here since it absolutely doesn‘t matter in this case where or why it happens...

 

[WithTheBirds]

One last example for binning on the chip. Different sensor but also Sony Exmor, so CMOS:

https://www.sony-semicon.co.jp/products_en/IS/sensor1/img/products/ProductBrief_IMX350_20171109.pdf

„pixel BINNING readout funtion“ - don‘t tell me it doesn’t happen on the chip ;-)!!

Over and out here since it absolutely doesn‘t matter in this case where or why it happens...

Let’s see if you can stay out- your the one prattling on... I understand what the difference is. It is the principal reason CCD maintained superiority over CMOS in low light noise performance for so long.

 

[Cliff_622]

Let’s see if you can stay out- your the one prattling on... I understand what the difference is. It is the principal reason CCD maintained superiority over CMOS in low light noise performance for so long.

I mean....yes, this is an interesting idea. Is DJI using Sony's binning readout or are they using Ambarella's pixel binning function? We have established that its possible to get binning from both of these places. (On chip or externally as well)

So....the much bigger question is;

Why DJI? Why are you doing this from either location?

Cliff Totten

 

[dwallersv]

Very true. Photos are equal to the P4P. It's all about the video processing. Having a 1inch-type sensor is only half the story. HOW YOU "READ" that sensor is the other half of the story.

First, I want to join the chorus of Hosanna's for doing this. Independent, real data is useful; people's subjective impressions can always be argued indefinitely.

Cliff, I really appreciate the technical details. I'm an EE, so it was all understandable to me. However, I'm also as naive as anyone about these functional/algorithmic details of image capture.

So, here's the salient point for me, and probably many others: We never imagined there was a second "half to the story". We quite naturally think that if you have a 20Mpx sensor, and advertise it as a feature, then 20M pixels are being captured and processed every still/frame.

This is just the natural intrpretation to have. While I doubt anyone could successfully sue for fraud, this sure feels like it to me.

 

[WithTheBirds]

I mean....yes, this is an interesting idea. Is DJI using Sony's binning readout or are they using Ambarella's pixel binning function? We have established that its possible to get binning from both of these places. (On chip or externally as well)

So....the much bigger question is;

Why DJI? Why are you doing this from either location?

Cliff Totten

Your clearly a person who is more interested In the technical nuances than the marketing department spin. When I was talking on chip I literally meant the sensor die. That’s where I started- CCD can combine pixels on chip prior to ADC. The various CMOS schemes might be implemented in the sensor package (usually are of course).

Our only hope for the current issue might be a firmware flash- whether this might be possible is largely determined by which SOC is used.

Question- are we sure that the P4P isn’t looking sharper straight off the SD card due to some form of sharpening being employed? What is seeming apparent is that the M2 footage isn’t as far behind once both are tweaked in post. The SOC is doing a lot in the M2- this is immediately obvious when you compare the video off the SD card with what is broadcast over the FPV downlink. The downlink has very obvious barrel distortion correction applied. Could this be as simple as an engineering decision where higher priority was afforded to optimising the occusync processing pipeline? Is this simply a decision to provide a more raw video file to give greater options in post (less sharpening applied)? I know when I acquired my first decent DSLR (the 1DS MkIII was king at the time) I initially thought the raw files were inferior to lesser bodies untill I understood how to work with them in post.

 

[dwallersv]

My theory is different, to me it is obvious the current M2P FW is far from being finalised, and DJI Marketing wanted to launch in August 23th

Dude, I was with you until here.

How do I pronounce "23th"? got stuck and couldn't read any further.

 

[kilomikebravo]

Perhaps he meant the 23st?

 

[Cliff_622]

Your clearly a person who is more interested In the technical nuances than the marketing department spin. When I was talking on chip I literally meant the sensor die. That’s where I started- CCD can combine pixels on chip prior to ADC. The various CMOS schemes might be implemented in the sensor package (usually are of course).

Our only hope for the current issue might be a firmware flash- whether this might be possible is largely determined by which SOC is used.

Question- are we sure that the P4P isn’t looking sharper straight off the SD card due to some form of sharpening being employed? What is seeming apparent is that the M2 footage isn’t as far behind once both are tweaked in post. The SOC is doing a lot in the M2- this is immediately obvious when you compare the video off the SD card with what is broadcast over the FPV downlink. The downlink has very obvious barrel distortion correction applied. Could this be as simple as an engineering decision where higher priority was afforded to optimising the occusync processing pipeline? Is this simply a decision to provide a more raw video file to give greater options in post (less sharpening applied)? I know when I acquired my first decent DSLR (the 1DS MkIII was king at the time) I initially thought the raw files were inferior to lesser bodies untill I understood how to work with them in post.

So there is certainly a sharpening element involved in these drone tests. But, sharpening and detail are two VERY different subjects. Sharpening is nothing more than "micro-contrast" and "pixel haloing". Sharpening does not add one pixel more of detail or one drop more resolution. In other words, you can sharpen a DVD until the cows come home but you can never make it look like a Blu-ray or, of course, 4k. 4k is a much higher frequency spatial sampling. There is a lot more "information" inside of it. As a wave form, it captures MUCH higher frequencies and has a much higher maximum Nyquist frequency threshold. Sharpening and more sharpening and more sharpening.... can achieve NONE of that.

+1 sharpening was done on every drone in my test because on Ambarella image processors, +1 pulls back on Ambarella's HORRIBLE, temporal noise reduction algorithm. Their noise reduction is very "generic" and not custom tailored to Sony sensors the way Sony's own BIONZ X processors are. Ambarella eats and squashes and flattens progressively harder and harder any tones below middle grey. It starts there and just gets worse and worse the closer you get to your shadows. The Ambarella A9 was the worst NR I have EVER seen to this date.

Anyway...if you want to capture more resolution, have less detail destruction, use +1. Yes, you WILL get more noise but you can remove that noise FAR....FAR.......FAR more surgically in post than any Ambarella noise reduction can do on the fly in camera. (and retain MUCH more detail while doing it)

Cliff Totten

 

[Jixxer]

Last reply from dji support concerning YouTube video

 

[Cliff_622]

I did see this a few days ago. It's funny,...it actually raises more questions than it answers. (it actually answers nothing....like "why" this was needed)

1.) Pixel binning IS EXACTLY in the category of "raw sub-sampling". Technically, so is line skipping. So DJI does admit that it is "sub-sampling" and that is why we see the artifacts of "sub-sampling". (Fro fun,....let's call it "DJI fancy binning"!)
2.) "Why" does FOV mode "REQUIRE" sub-sampling? (DJI Fancy Binning") Why?
3.) "How" is DJI's "raw sub-sampling" more "advanced" than simple pixel binning? The end result is IDENTICAL to "simple" sub sampling..i.e..pixel binning!
4.) DJI's "raw subsampling" is not visible on real scenes? (only on charts?) What?....total BS!! It's immediately visible to the naked eye.

They told us nothing with this. They say; "we do a special raw sub-sampling" (exactly what pixel binning is) but ours is somehow better and cant be noticed....even though the end result is no different than "simple" pixel binning and it is REQUIRED on the Mavic-2 Pro in FOV.

C'mon guys,...I love you, I really do!...but you cant dance around this. You just spoke about 3 paragraphs and said exactly NOTHING that we already haven't figured out before hand.

Cliff Totten

 

[jmarka]

Cliff - do you know if this line skipping is also true for still image photography? My suspicion is it is skipping there too. My m2p images seem no higher resolution than my original mp

 

[Cliff_622]

Cliff - do you know if this line skipping is also true for still image photography? My suspicion is it is skipping there too. My m2p images seem no higher resolution than my original mp

Nah,...photography on this is solid. This is only a 4k FOV video thing.

 

[WithTheBirds]

ma

I did see this a few days ago. It's funny,...it actually raises more questions than it answers. (it actually answers nothing....like "why" this was needed)

1.) Pixel binning IS EXACTLY in the category of "raw sub-sampling". Technically, so is line skipping. So DJI does admit that it is "sub-sampling" and that is why we see the artifacts of "sub-sampling". (Fro fun,....let's call it "DJI fancy binning"!)
2.) "Why" does FOV mode "REQUIRE" sub-sampling? (DJI Fancy Binning") Why?
3.) "How" is DJI's "raw sub-sampling" more "advanced" than simple pixel binning? The end result is IDENTICAL to "simple" sub sampling..i.e..pixel binning!
4.) DJI's "raw subsampling" is not visible on real scenes? (only on charts?) What?....total BS!! It's immediately visible to the naked eye.

They told us nothing with this. They say; "we do a special raw sub-sampling" (exactly what pixel binning is) but ours is somehow better and cant be noticed....even though the end result is no different than "simple" pixel binning and it is REQUIRED on the Mavic-2 Pro in FOV.

C'mon guys,...I love you, I really do!...but you cant dance around this. You just spoke about 3 paragraphs and said exactly NOTHING that we already haven't figured out before hand.

Cliff Totten

The explanation seems reasonable and acceptable. How might you propose they might get to 4K frame width absent some form of interpolation (assuming the whole sensor is employed across the frame)?

Downsizing does increase detail (for a given area smaller than that prescribed by the native pixel dimension) but you loose sharoness. Interpolation causes a loss of contrast between adjoining pixels. This can and will be effectively recovered by applying an unsharp mask filter.

No more worrying about binning or overheating theories.

 

[chunky]

Let’s see if you can stay out- your the one prattling on... I understand what the difference is. It is the principal reason CCD maintained superiority over CMOS in low light noise performance for so long.

So it proves I‘m more on the facts, seems you are still one way ;-)...
Nothing to struggle here, at least it doesn‘t matter where what ever happens, we only can count on the specs, not on any technique someone is using for what case ever...

 

[Lon Denard]

Nah,...photography on this is solid. This is only a 4k FOV video thing.

Man, I really, really hate you! Hahaha! I've been fighting with people all day because of you. You've given all the DJI haters fuel to use in every YouTube video to come out since yours.

DJI never claimed you would get P4P video with the Mavic 2 Pro. It's like they marketed it as the folding P4P or something. They promised the M2P would be an improvement over the MP, it is. Not that it would as good or better than the P4P, it isn't. Besides, to say that they are hiding the fact is ridiculous, all of the specs are on the website for all to see. If you place them next to the P4P specs, they aren't the same.

As far as the accusation that DJI hyped the drone beyond what it would deliver. Nah that was YouTube reviewers. A day couldn't go by without one or two of them blathering on about how awesome the M2P was going to be. DJI had their release, went over the specs but I never heard him once claim that it was as good as the P4P. That was the YouTube hype train promoting that idea.

I must have missed the video complaining that the Spark uses the same sensor as the Pro, why doesn't it shoot in 4k?

I'm just waiting for Bo to do a review. If he says the camera is trash, then fine I'll throw the thing away and go out and buy an Evo and be done with it.

 

[chunky]

I mean....yes, this is an interesting idea. Is DJI using Sony's binning readout or are they using Ambarella's pixel binning function? We have established that its possible to get binning from both of these places. (On chip or externally as well)

So....the much bigger question is;

Why DJI? Why are you doing this from either location?

Cliff Totten

I guess it‘s only possible on the sensor here because of speed/performance issues. Getting 30p in 16:9 full readout is „by far“ not specified (only up to 26p) and so this may be the only chance!
Since I‘m PAL 25p or 24p would also be ok for me in full readout.

BTW: we know nothing, not if it‘s an Ambarella SoC nor if it‘ an IMX183 nor if it‘s a Sony Sensor!
On the P4p we know we have Ambarella H1 and DJI states themselves we have a Sony Sensor.

Here we only are promised we get 4K videoresolution (!!) also in FullFOV

 

[chunky]

So there is certainly a sharpening element involved in these drone tests. But, sharpening and detail are two VERY different subjects. Sharpening is nothing more than "micro-contrast" and "pixel haloing". Sharpening does not add one pixel more of detail or one drop more resolution. In other words, you can sharpen a DVD until the cows come home but you can never make it look like a Blu-ray or, of course, 4k. 4k is a much higher frequency spatial sampling. There is a lot more "information" inside of it. As a wave form, it captures MUCH higher frequencies and has a much higher maximum Nyquist frequency threshold. Sharpening and more sharpening and more sharpening.... can achieve NONE of that.

+1 sharpening was done on every drone in my test because on Ambarella image processors, +1 pulls back on Ambarella's HORRIBLE, temporal noise reduction algorithm. Their noise reduction is very "generic" and not custom tailored to Sony sensors the way Sony's own BIONZ X processors are. Ambarella eats and squashes and flattens progressively harder and harder any tones below middle grey. It starts there and just gets worse and worse the closer you get to your shadows. The Ambarella A9 was the worst NR I have EVER seen to this date.

Anyway...if you want to capture more resolution, have less detail destruction, use +1. Yes, you WILL get more noise but you can remove that noise FAR....FAR.......FAR more surgically in post than any Ambarella noise reduction can do on the fly in camera. (and retain MUCH more detail while doing it)

Cliff Totten

+1 for part 1, but I wouldn‘t say sharpening helps here really. I only found noise kicking in quite early and getting rid of noise with Neat Video is still lot‘s of computing power, even with stromg GPUs.
Resolve Studio is not the same league.