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

[chunky]

As just posted in the RC-Groups I am starting to think we really have "FOV-Gate" developing. In my test chart pics, it is so obvious that the artefacts in the details in 2,7K and 4K FOV are nearly identical. Having 5472×3648px I presume they do the following:

Take the full sensor (cropped to 16:9)

4K = full sensor with 2x2 pixel binning resulting in 2,7K and let the SOC inflate it to 4K
2,7K full sensor with 2x2 pixel binning resulting in 2,7K
1080 full sensor with 4x4 pixel binning resulting in 1080

This should be extremely fast, quick and dirty to implement and would explain my test results and provide the reason for 2,7K and 1080p being in FOV by default..
What do you guys think?

I‘m with you here!
4K FFOV shows the same resolution and the same artifacts as 2.7K30p where 4K „profits“ a bit from some additional interpolation, looks a bit less pixelated but in the end shows not more detail.
2.7K60p shows even more artifacts than 2.7K30p, looks even less usable.


Last question I see is how does a synthetic ISO chart downscaled to 2.7K look like?!!
How does „real“ 2.7K compare to M2P 2.7K??

 

[mozso]

As just posted in the RC-Groups I am starting to think we really have "FOV-Gate" developing. In my test chart pics, it is so obvious that the artefacts in the details in 2,7K and 4K FOV are nearly identical. Having 5472×3648px I presume they do the following:

Take the full sensor (cropped to 16:9)

4K = full sensor with 2x2 pixel binning resulting in 2,7K and let the SOC inflate it to 4K
2,7K full sensor with 2x2 pixel binning resulting in 2,7K
1080 full sensor with 4x4 pixel binning resulting in 1080

This should be extremely fast, quick and dirty to implement and would explain my test results and provide the reason for 2,7K and 1080p being in FOV by default..
What do you guys think?

This was my worst nightmare scenario. I cannot test or contribute as I suspended my M2 Pro purchase.

I was hoping for:
- 1080p resolution with 2.7K sensor readout and proper downsampling
- 2.7K resolution with 4K sensor readout and proper downsampling
- 4K resolution with 5.5-5.7K sensor readout and proper downsampling

By proper downsampling I mean downscaling, not line skipping or pixel binning.

Without these the marketed 1” sensor is just a marketing trick, nothing more. Yes its better by default than the old smaller sensor in noise and others, but its full potential is not leveraged. As proven by P4Pro V2 comparisons.

Question is:
- will DJI correct this by firmware fix (product launch rush forced them to skip proper implementation)
- this is by design as is and you need to buy a V2/Platinum/Enterprise edition of M2Pro to have it fixed

 

[chunky]

DJI double screwed customers. For early adopters (like me) who purchased on announcement day, we were promised "superior 10bit 4K" powered by Hasselblad, and we got what we got, the crappy 4K FOV.

Then for those who waited a bit, now both the M2P and the Fly More Combo are an extra 110 euro combined, and the issues are known making it tougher to make the purchase.

I really hope sales take a hit and they're forced to fix 4K FOV. Just use the algorithms used in the P4P, the hardware is the same or better (the SoC). As of today 4K FOV it is basically 2.7K FOV resized.

True, really some kind of ripoff!!

But back to the SoC - is it really what some people „guess“... Ambarella H3??
„I“ have no real prove as with the P4Pro where we have (!) an Ambarella H1 (NOT H2).
Does this really help us if we know what SoC is used?

 

[endless_shutters]

If these theories prove to be factual, what action can we take as consumers?

 

[Sockeye]

But back to the SoC - is it really what some people „guess“... Ambarella H3??

The DJI Supporter in the RC group stated, that folks who speculate that the SOC is an Ambarella, are 100% wrong..

 

[Sockeye]

If these theories prove to be factual, what action can we take as consumers?

I am a film maker, not a lawyer... and pissed because I let DJI Marketing manage my expectations again...

 

[chunky]

Gentlemen, fasten seatbelts - this is what 2.7K synthetic resolution would be able to resolve
(scale is same as for 4K, so 2,81% for the centerchart of the 16:9 picturewidth, whole testchart equals 50% of the cameras 16:9 picturewidth):

Seems debayering and „some other factors“ „eat“ a significant additional amount of resolution.

The M2P resulting resolution seems only to equal ideal FullHD (like from a 3chip or monochrome camera).

 

[Sockeye]

I am not sure on how you constructed the "synthetic 2.7K"?

 

[yysc]

The DJI Supporter in the RC group stated, that folks who speculate that the SOC is an Ambarella, are 100% wrong..

He's been wrong in the past. Just a support guy (not very knowledgeable in video) who is fed limited information from Engineering and seems to have limited access to DJI internal product development. I wouldn't take his word as final DJI stance.

 

[chunky]

I am not sure on how you constructed the "synthetic 2.7K"?

Took the ISO12233 testchart as vector .pdf and imported it to Photoshop in a „very high“ resolution.
Cropped it to the 16:9 marks of the chart and doubled the picturewidth (no upscale, just added blank space).
Then did a bicubic downscale to 2688 horizontal pixels (2.7K) and cut out the upper crop.
(In real I have downscaled the 16:9 crop of the chart to 1344 horizontal pixels what is the same, but you don‘t see „2.7K“ here).
For comparison with your crops I upscaled it by near neighbor to 600%.

 

[chunky]

Maybe you all should make screenshots of the currents „specs“ and download a copy of the 1.0 user manual for reference where 4K „resolution“ is stated.
If it‘s right a new version of the manual will come tomorrow.

 

[WithTheBirds]

Many of the current sensors, especially Sony Exmor can do the binning on the sensor itself (Exmor also contains the ADC).

Here is a nice explanation for Quad HDR adaptive binning on a Sony chip that also happens on the sensor before readout and further processing in the SoC:
IMX294CJK | Sony Semiconductor Solutions

There was also a nice explanation for on sensor binning I think on FRAMOS but I don’t find it anymore.

Whatever happened, the current result for the FullFOV mode is a whatever flaw and the explanation of the potential DJI guy is it even more.
I can do nothing with „RAW subsampling“, for
me it sounds like binning that is even an advantage over line skipping because of the better SNR.
If „RAW subsampling“ is simple on sensor binning for saving readout efforts then it‘s simply „binning“ to 2.7K with no excuse - bad since the P4pro was capable of much more.
If they read the full sensor and do their „RAW subsampling“ in their „processor“ - then it‘s unclear why this „quality reduction“ happens anyway since every needed information is already present in the pipe.

We can break our heads but actually it is like it is, we bought a 4K drone with only the power of 2.7K in FullFOV.

Even worse it seems the „native“ 2.7K modes become even less powerful in fps‘es over 30p (haven‘t checked that myself yet).

Its not greatly important to the discussion here however you seem really intent on argument- CMOS sensors do not perform true binning, they average the outputs of individual sensor photo sites. Think of each photo site as a bucket and the light photons as golf balls, peanuts- whatever you like. True binning (which no CMOS sensor can perform) tips buckets into each other before counting. The process provides a genuine increase in SNR.

 

[chunky]

As long as Sony themselves as manufacturer call it „binning“ I think it‘s quite ok to speak about binning ;-).
It is technically not the same as it is on CCDs, but it‘s done on the chip and means that „several“ pixels are logically „coupled“ to a common „cluster“.

 

[WithTheBirds]

As long as Sony themselves as manufacturer call it „binning“ I think it‘s quite ok to speak about binning ;-).
It is technically not the same as it is on CCDs, but it‘s done on the chip and means that „several“ pixels are logically „coupled“ to a common „cluster“.

It’s not the same- I suspect you worked that out quickly. The data sheet you posted is for a security camera CMOS- in some ways we might be better with that device as it has a lot more all pixel scan read out modes. What we call if doesn’t really matter as Interesting topic as it is (fascinating) let’s keep this relavent and look at what we might be dealing with. Vertical and/or horizontal line addition (as per Sony information) at the sensor level and who knows what else in the SOC is our issue. I any case assuming the M2P is running the Exmor R IMX183CQJ 4096 *2160 is available in 10bit "All Pixel Scan" (i.e no "binning" or need to implement line skipping, adding or anything).

All I know is I’m looking at the actual user findings to date where your testing is especially informative and shaking my head. Something is def not right- lets hope it’s simply an issue of dodgy firmware which DJI are working on a fix for.

 

[Cliff_622]

The 5 min limit is SONY's own design issue, Panasonic LX10 (similar in size to RX100) also uses SONY 1" BSI 20MP sensor and such limitation does not exist in 4K.

Also GH5 using SONY sensor does not overheat and many of the competitor SONY cameras do.

Absolutely. And generally, these 1inch-type sensors run somewhat cool. I run a time hack on my RX10-III and have made 4 hour recordings with it and had no heat problems.

Now,..these are all still much larger cameras and the Do get warm.

I cant say this with 100% confidence,...but the Mavic-2 Pro is by far the smallest 1inch-type camera I have know of and it might be THE smallest 1inch-type sensor camera ever made by anybody.

That sensor truly is PACKED into a very tine cube!

Again,...only DJI knows the real reason why they chose to read that sensor in FOV in the way they did.....a REALLY ODD choice!!

CT

 

[Vegasrobbi]

Want to give you guys a thumbs up to for discussing this and looking for answers to make the M2P better. Try this over at RCG and you end up in the hole.

Don't trust the reps over there they know things their not telling and telling things their not knowing.

 

[WithTheBirds]

Want to give you guys a thumbs up to for discussing this and looking for answers to make the M2P better. Try this over at RCG and you end up in the hole.

Don't trust the reps over there they know things their not telling and tell things their not knowing.

No point knowing and not sharing.....

 

[chunky]

It’s not the same- I suspect you worked that out quickly. The data sheet you posted is for a security camera CMOS- in some ways we might be better with that device as it has a lot more all pixel scan read out modes. What we call if doesn’t really matter as Interesting topic as it is (fascinating) let’s keep this relavent and look at what we might be dealing with. Vertical and/or horizontal line addition (as per Sony information) at the sensor level and who knows what else in the SOC is our issue. I any case assuming the M2P is running the Exmor R IMX183CQJ 4096 *2160 is available in 10bit "All Pixel Scan" (i.e no "binning" or need to implement line skipping, adding or anything).

All I know is I’m looking at the actual user findings to date where your testing is especially informative and shaking my head. Something is def not right- lets hope it’s simply an issue of dodgy firmware which DJI are working on a fix for.

See table 3 in the IMX183 product info:
...
*1 With horizontal addition
*2 With vertical addition

That’s on chip BINNING Sony call‘s it themselves.

 

[chunky]

Want to give you guys a thumbs up to for discussing this and looking for answers to make the M2P better. Try this over at RCG and you end up in the hole.

Don't trust the reps over there they know things their not telling and telling things their not knowing.

Thumbs up to all those who also ask the DJI support „WHEN“ the update appears from where the M2P will fulfill the specified (!!) video „resolutions“.

 

[WithTheBirds]

See table 3 in the IMX183 product info:
...
*1 With horizontal addition
*2 With vertical addition

That’s on chip BINNING Sony call‘s it themselves.

Why? Seriously.... call it whatever you like. I already posted the table you refer to when I first made the point. Binning is combining photo sites on the sensor to create a larger “bin”.... effectively the summing is done in the analogue domain prior to AD conversion. You get at a true increase in SNR. I’m sorry you feel you have to keep labouring this point... I’m happy to say every technique that arrives at a lower resolution without changing FOV is binning...

Have a good look at the table and you will see there are only two modes where horizontal and/or vertical binning isn’t performed. It might be a clue in solving this mystery. That was my intent in originally posting it although I almost wish I hadn’t.

While your at it another observation might be the white paper says nothing about line skipping.

I suspect we will ultimately find sensor heating issues are not the issue here.