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

[Lon Denard]

My conclusions so far:

4k FOV is not just upsampled 2.7k, but that would be a ridiculously poor strategy - I can't believe that anyone really thought that.​

​

4k FOV is worse than the M2P 20 MP images downsampled to 4k, but that was expected - we just don't know how they are being generated.​

​

4k FOV is not too dissimilar to 4k on the Mavic Pro, but since both have more than enough resolution for 4k, that's not particularly surprising. The 10 bit output from the M2P is superior by other metrics.​

​

4k HQ is almost exactly as expected in compared to the full 20 MP images.​

​

None of the results quite lives up to the theoretical maximum shown above, but given the additional factors of real lens effects, low-pass filters etc., it's credibly close.

Well that's cool! I'm still sticking with HQ though. I like the close up feel to it.

 

[Lon Denard]

Thank you for your efforts here SAR- no doubt many will find the results very informative.

It would seem that we can now safely almost completely discount certain of the previous assumptions, particularly the suggestion that the sensor performance was crippled by heat issues or that the 4K FOV output was simply an upsampled 2.7K sensor read.

It is doubtful DJI has tried to pull as swifty on us.

If in fact DJI employed the Sony IMX183CQJ sensor in the M2P a quick perusal of the published specifications reveals that the only readout mode that supports all pixel scan in 10bit depth above 25fps is "type 1/1.4" which provides 9.03mpx max at 17:9 aspect ratio (reading 71% of the sensor horizontally). All other readout modes above 25fps employ a combination of horizontal and/or vertical line addition.

FOV mode must be subject to the sensor limitations, that is line addition is, and must be, employed prior to further SOC processing.

DJI has perhaps shot themselves in the foot by not providing further explanation as the the technical issues with the camera and the design decisions. HQ mode looks like a genuine attempt to give us something extra. Unfortunately they are being criticised by those of us who might not appreciate they have done the best they might have done with FOV mode given the Sensor limitations.

If I end up being right all along, I'm gonna laugh myself silly!!!

 

[WithTheBirds]

If I end up being right all along, I'm gonna laugh myself silly!!!

Try not to get any broken ribs in the immediate future.

 

[Lon Denard]

Try not to get any broken ribs in the immediate future.

I'll try...

 

[WithTheBirds]

I'll try...

Only thing worse than laughing with broken ribs is sneezing...

 

[Lon Denard]

What I was delicately trying to suggest you was - not to use Mlog. It is much to much unsharp (unclear). Also, you'll not get any more detail from it. Not to mention the noise. Just be patient and wait for next FW.
In the meantime you can do colorgrade from vid. in "normal" color, using all HQ benefits (except 10 bit).
It was meant to be helpful.

p.s.
I used this clip just to show how much aliasing you can expect in hq

Btw you don't live in Munich do you? I ordered a case from MC Cases and I believe that's where they're made. Give them a kick in the butt and make them get my case to me! Oh and if you go to Stuttgart, stop by Paradise and say hi the Petra for me. Blonde Petra not brunette or red headed Petra.
Thanks.

 

[Dronefriend]

Are you saying this problem only exists in 30p NTSC and not PAL or 24p?

 

[ABLomas]

The second and third images show the theoretical maximum resolutions available at 4k with this image.

Uh, no. If you are comparing line count - then do not use bicubic resampling for that (if you want to show maximum resolution). Use at least something like lanczos or spline - which keeps sharp lines and details. Bicubic is "smooth" and soft - it's great in most cases, but not for pixel peeping. Yes, you will get more aliased lines, but resolving power should be much, much higher.
In-camera resizing is probably something like bilinear (or, heh, in this case looks more like "nearest" to me) so again, not so smooth as bicubic.

Funny to see few meaningfull posts here and suddenly poof! again, entire page of useless chit-chat offtopic from ppl who thinks that they MUST answer to every post in this thread.

 

[Lon Denard]

Uh, no. If you are comparing line count - then do not use bicubic resampling for that (if you want to show maximum resolution). Use at least something like lanczos or spline - which keeps sharp lines and details. Bicubic is "smooth" and soft - it's great in most cases, but not for pixel peeping. Yes, you will get more aliased lines, but resolving power should be much, much higher.
In-camera resizing is probably something like bilinear (or, heh, in this case looks more like "nearest" to me) so again, not so smooth as bicubic.

Funny to see few meaningfull posts here and suddenly poof! again, entire page of useless chit-chat offtopic from ppl who thinks that they MUST answer to every post in this thread.

Why not answer every post? Why not joke around a little bit? What's up with the thread police for goodness sake? Yeesh. Fellas these are TOYS at the end of the day. They aren't even that expensive. I like joking around and acting all serious but in the end, it's just not that critical. Un follow the thread if it bothers you. Put me and whomever else you want on "ignore" since I don't post anything of merit. This is not the American Journal of Medicine forum and we aren't discussing cancer treatments or pacemakers.

 

[sar104]

Uh, no. If you are comparing line count - then do not use bicubic resampling for that (if you want to show maximum resolution). Use at least something like lanczos or spline - which keeps sharp lines and details. Bicubic is "smooth" and soft - it's great in most cases, but not for pixel peeping. Yes, you will get more aliased lines, but resolving power should be much, much higher.
In-camera resizing is probably something like bilinear (or, heh, in this case looks more like "nearest" to me) so again, not so smooth as bicubic.

Funny to see few meaningfull posts here and suddenly poof! again, entire page of useless chit-chat offtopic from ppl who thinks that they MUST answer to every post in this thread.

I used bicubic because the entire preceding discussion had focussed on bicubic, and I didn't want to change that variable as well.

 

[Dronefriend]

Guys, simple and very important question for everyone located outside of the States.

Is the problem only in 30fps mode? Do you get full read out in 25p/24p?

 

[sar104]

Uh, no. If you are comparing line count - then do not use bicubic resampling for that (if you want to show maximum resolution). Use at least something like lanczos or spline - which keeps sharp lines and details. Bicubic is "smooth" and soft - it's great in most cases, but not for pixel peeping. Yes, you will get more aliased lines, but resolving power should be much, much higher.
In-camera resizing is probably something like bilinear (or, heh, in this case looks more like "nearest" to me) so again, not so smooth as bicubic.

Funny to see few meaningfull posts here and suddenly poof! again, entire page of useless chit-chat offtopic from ppl who thinks that they MUST answer to every post in this thread.

Take your pick. Lanczos increases sharpening but at the expense of more artifacts. It doesn't change the relative outcome - it just makes it uglier.



High-resolution crop


Bicubic to 20 MP equivalent


Lanczos 3 to 20 MP equivalent


Lanczos 3 (non-separable) to 20 MP equivalent


Downsampled to 2.7k then upsampled to 4k (bicubic)


Downsampled to 2.7k then upsampled to 4k (Lanczos 3 non-separable)

 

[Lon Denard]

Guys, simple and very important question for everyone located outside of the States.

Is the problem only in 30fps mode? Do you get full read out in 25p/24p?

That would also be crazy, if true and would explain why some people are adamant that the video is horrible and some say that it's fine. I doubt it but it's something to think about for sure.

 

[ABLomas]

Is the problem only in 30fps mode? Do you get full read out in 25p/24p?

Def. no, at least on mine it's visible on 25 and 30FPS.
I did not tried 24 FPS.

Take your pick. Lanczos increases sharpening but at the expense of more artifacts. It doesn't change the relative outcome - it just makes it uglier.

Thank you.
Strange results, actually, i expected much cleaner output (esp. on non-straight lines, because bicubic would interpolate everything around and everything should become gray mess; expected that lanczos would eliminate that aliasing artefacts and keep white/black areas), but anyway, i believe that you done everything correctly.

Why not answer every post? Why not joke around a little bit?

Community Guidelines
#4, #13 and so on...

 

[sar104]

Thank you.
Strange results, actually, i expected much cleaner output (esp. on non-straight lines, because bicubic would interpolate everything around and everything should become gray mess; expected that lanczos would eliminate that aliasing artefacts and keep white/black areas), but anyway, i believe that you done everything correctly.

The angled line pairs seem to be very tough tests of optical systems and the resample and sharpen algorithms. With real images I've given up trying to predict which will give the best results, especially since the aliasing issue means that sharpest is often not the best.

 

[Lon Denard]

Community Guidelines
#4, #13 and so on...

So, another member replied about broken ribs, I answered and that's a problem? How about you let the actual mods moderate and YOU can stick to the topic at hand? Trust me, I'm not allowed to get too far afield around here. If there were a problem, it would be handled immediately.

 

[yysc]

That would also be crazy, if true and would explain why some people are adamant that the video is horrible and some say that it's fine. I doubt it but it's something to think about for sure.

The 4K FOV video is not horrible but:

1. It uses an inferior downsampling algorithm (subsampling) as opposed to that (supersampling) used in the other DJI 1" sensor consumer drones (P4P/P4A). Therefore there is less detail that in the aforementioned drones using similar HW.
2. As a result of the subsampling approach in 4K FOV there are aliasing issues which are not present to this degree in other 4K drones. These cannot be fixed in post.

sar104 is trying to stablish 4K FOV is similar in terms of resolving detail to 4K in the M1P for instance, or that there is more detail in 4K than 2.7K in the M2P, and while that is fine it does not negate the two issues mentioned above.

M2P exports to both H264 8 bit (like the P4P) and H265 10 bit. I can't see why they cannot use the same downsampling algorithm in the 8 bit mode and give users the choice if there is a limitation (in the sensor readout in 10bit mode), considering the SONY sensor is likely the same.

 

[sar104]

The 4K FOV video is not horrible but:

1. It uses an inferior downsampling algorithm (subsampling) as opposed to that (supersampling) used in the other DJI 1" sensor consumer drones (P4P/P4A). Therefore there is less detail that in the aforementioned drones using similar HW.
2. As a result of the subsampling approach in 4K FOV there are aliasing issues which are not present to this degree in other 4K drones. These cannot be fixed in post.

sar104 is trying to stablish 4K FOV is similar in terms of resolving detail to 4K in the M1P for instance, or that there is more detail in 4K than 2.7K in the M2P, and while that is fine it does not negate the two issues mentioned above.

M2P exports to both H264 8 bit (like the P4P) and H265 10 bit. I can't see why they cannot use the same downsampling algorithm in the 8 bit mode and give users the choice if there is a limitation (in the sensor readout in 10bit mode), considering the SONY sensor is likely the same.

"sar104 is trying to stablish 4K FOV is similar in terms of resolving detail to 4K in the M1P for instance, or that there is more detail in 4K than 2.7K in the M2P, and while that is fine it does not negate the two issues mentioned above."​

I'm not trying to establish anything - I've simply made some measurements to address unsupported assertions and prior experiments that seemed lacking in rigor.

Your question about using the same method as the P4P in 8-bit mode is fair, but even the premise is speculation given that we don't actually know which sensor and which image processor is in the M2P.

 

[yysc]

"sar104 is trying to stablish 4K FOV is similar in terms of resolving detail to 4K in the M1P for instance, or that there is more detail in 4K than 2.7K in the M2P, and while that is fine it does not negate the two issues mentioned above."​

I'm not trying to establish anything - I've simply made some measurements to address unsupported assertions and prior experiments that seemed lacking in rigor.

Yes,

My conclusions so far:

4k FOV is not too dissimilar to 4k on the Mavic Pro, but since both have more than enough resolution for 4k, that's not particularly surprising. The 10 bit output from the M2P is superior by other metrics.​

​

​

Although by your very own assertion it is certainly very underwhelming, that a much more expensive (50% price increase) and newer drone, with a much better sensor, cannot resolve more detail that the older precedessor. And yes, the 10 bit color is nicer, but that's only a part of the overall IQ.

Your question about using the same method as the P4P in 8-bit mode is fair, but even the premise is speculation given that we don't actually know which sensor and which image processor is in the M2P.

Right, but considering a cheap ($300 or less) action camera like the Yi 4K+, released in Q1 2017, already uses a far better SOC than DJI in the P4P (Amba H2 vs H1), it would be extremelly embarrassing if they used something less capable. And less capable than the H1 is the junk A9 used in the original M1P.

That certainly would cost peanuts, but we know this is not the case, as the M2P resolves shadow detail better than the H1 in the P4P meaning it has a SOC with superior NR algorithms than the H1. It would very very weird that this newer SOC can do that and not a downsample as the H1 which is nowadays obsolete.

 

[OndaEyes]

Put me and whomever else you want on "ignore" since I don't post anything of merit.

Thanks for the hint !