Is 22.3MP the magic number for 1080 video from Canon?

[kye]

I have a theory about why Canon video quality is typically bad, and which models may be better than others.

We know from Magic Lantern that Canon uses every third pixel across the sensor to get 1920 x 1080.  The problem is that for most of their cameras there aren't quite enough pixels.

For my 18MP 700D, which is 5184x3456 that means the 3x3 resolution is 1728x1152 which cropped to 16:9 is 1728x972.  This then needs to be upscaled 111.11% to get 1920.  
We all know that upscaling video is not a nice thing to do, so despite Canon probably applying sharpening before compressing it, the recipe of RAW video -> upscaled -> processed -> compressed isn't a recipe for success!  
I suspect the combination of upscaling combined with the heavy ~56Mbps compression is the culprit as upscaling tends to soften detail and compression tends to crunch things that aren't sharp.  We know that sharp images can still survive a 50Mbps codec, and my ML RAW experiments seem to indicate that 1728x1152 isn't fundamentally terrible if treated nicely.

This leads me to question what the right recipe is.
If we start with 1920 and work backwards, we get 1920*3 = 5760.  This is how wide the sensor has to be for a 3x3 reduction to not need upscaling.  If the sensor is 3:2 then we need 5760x3840 which is 22.21MP.

Therefore, my theory is that all the Canon cameras with resolutions above 22.3MP should have superior 1080p quality.

According to this comparison table this would mean that the 5DIII, 5D4, 5DS, 6DII, 77D, 80D, 750D, 760D, 800D, 2000D, 200D, M3, M5, M6, M50 and M100 are the potential winners.

Does people's experience of these cameras back this up?

Of course, if any of the <22.3MP cameras took a higher resolution reading of the sensor and downscaled it then they would produce a nice 1080 image, but I don't know if any Canon DSLRs work in this way?

[Inazuma]

All of those cameras have varying image quality unfortunately. Even some of the cameras with same sensors as each other have different video quality, such as the 80d and m6.

[Nikkor]

forget about canon

[Andrew Reid]

This is assuming they actually read out exactly 1920 x 1080 on the sensor.

It is possible by varying the resolution very slightly to do a 3x3 binning on almost any sensor.

The best Canon 1080p resolution is delivered on the 5DS funnily enough, at 50MP!

Also it could be that Canon read out a lower than 1920 x 1080 image on some of their cameras to purposefully limit things - either for power consumption or heat reasons or marketing. The Magic Lantern raw video frame sizes from a lot of the 1080p Canons prove this.

Sony have a very good 1080p (and 4K) image from their 42MP sensors.

Sony again manage an incredible 4K image from a 100MP Hasselblad sensor (see below!)

So there is not really any magic megapixel count, certainly not 22.3MP.

 

[kye]

Damn, was hoping those cameras with enough pixels would be one destructive processing step less (which they might well be) and that it would be one of the ones that did the most damage (which it sounds like it isn't).

[DBounce]

4 hours ago, Andrew Reid said:

This is assuming they actually read out exactly 1920 x 1080 on the sensor.

It is possible by varying the resolution very slightly to do a 3x3 binning on almost any sensor.

The best Canon 1080p resolution is delivered on the 5DS funnily enough, at 50MP!

Also it could be that Canon read out a lower than 1920 x 1080 image on some of their cameras to purposefully limit things - either for power consumption or heat reasons or marketing. The Magic Lantern raw video frame sizes from a lot of the 1080p Canons prove this.

Sony have a very good 1080p (and 4K) image from their 42MP sensors.

Sony again manage an incredible 4K image from a 100MP Hasselblad sensor (see below!)

So there is not really any magic megapixel count, certainly not 22.3MP.

 

That’s a nice looking image from the Hasselblad... but it’s got to be a “hassle”  to work with given it’s hybrid nature. At that price point ($33k), there are many better options imo.

[Django]

Canons best 1080p is on it's C line.  Even the lowly 24Mbps from the C100 blows away in sharpness/detail any Canon DSLR with twice the bitrate. That's cuz of internal downscaling from a 4k sensor.

Of course ML Raw on 5D3 is in another league but the typical Canon soft 1080p is terrible across the entire DSLR line imo and feels like upsacled 720p..

[kidzrevil]

They chose 22.3 because it allows them to do 3x3 pixel binning without moire artifacts or a noticeable degradation of quality. Pixel binning also has the added effect of making pixels look brighter by using neighboring pixels from the 22.3 sensor when output to 1080p. It cuts down on processing power required versus 1:1 full sensor readout and works really well for the higher end Canon's that don't use line skipping for their 1080p output

1 hour ago, Django said:

Canons best 1080p is on it's C line.  Even the lowly 24Mbps from the C100 blows away in sharpness/detail any Canon DSLR with twice the bitrate. That's cuz of internal downscaling from a 4k sensor.

Of course ML Raw on 5D3 is in another league but the typical Canon soft 1080p is terrible across the entire DSLR line imo and feels like upsacled 720p..

I think its the codec and in camera processing making the image so soft. ML Raw is such a huge jump in quality it has to be the codec & image processing 

[Django]

1 hour ago, kidzrevil said:

They chose 22.3 because it allows them to do 3x3 pixel binning without moire artifacts or a noticeable degradation of quality. Pixel binning also has the added effect of making pixels look brighter by using neighboring pixels from the 22.3 sensor when output to 1080p. It cuts down on processing power required versus 1:1 full sensor readout and works really well for the higher end Canon's that don't use line skipping for their 1080p output

I think its the codec and in camera processing making the image so soft. ML Raw is such a huge jump in quality it has to be the codec & image processing 

the codec isn't helping but it's more than that.. cuz recording to prores externally stills gives out a soft image. the upscaling seems to be the main culprit.

someone in this thread claimed: 

5D3 uses large group pixel bins of 5x5, resulting a 1152x648 RAW bayer, which then gets upscaled to 1080p, that's why it has about 600-700 lines of resolution

 

 

 

[kidzrevil]

@Django the Canon 5D mark iii is for sure 3x3 pixel binning. I recall reading an article from a canon engineer saying that they didn’t aim to increase the megapixels over the Canon 5D mark ii because with 3x3 pixel binning scales to 1080p without skipping lines. Pixel binning works by combining neighboring pixels and 22.3 with their 3x3 binning equates a 1080p 2mp image.

5760 x 3840 divided by 3x3 = 1920 x 1280 3:2 aspect ratio. The 200 pixels get cropped out for 1920 x 1080 video

What seems to be happening with Canon RAW video in magic lantern is bypassing a lot of the elements that equate the Canon soft resolution. 

[kidzrevil]

@Andrew Reid the 5DS has a LOT of line skipping to get to that 2 megapixel 1080p image. Im not sure I’ve seen or heard anything good come out of that camera especially in comparison to canon’s other dslr

[kye]

6 hours ago, Django said:

Canons best 1080p is on it's C line.  Even the lowly 24Mbps from the C100 blows away in sharpness/detail any Canon DSLR with twice the bitrate. That's cuz of internal downscaling from a 4k sensor.

Awesome, thanks..  So, 4K downscaling + 24Mbps = nice image.  This means that the 50Mbps bit-rate isn't fundamentally flawed.

I'd also suggest that YT videos are pretty low bitrate when they go from YT to the viewer so that's a second point of reference for bit-rate.

I'm not sure if all H264 encoding is of the same quality, so maybe there is some other limitation to the codec other than bitrate.

4 hours ago, kidzrevil said:

I think its the codec and in camera processing making the image so soft. ML Raw is such a huge jump in quality it has to be the codec & image processing 

Agreed.

I'm skeptical that there's a huge difference between H264 encoders (other than the bitrate) and a quick google didn't reveal lots of people comparing encoders so it doesn't appear to be a thing (happy to be proven wrong though!).  

This would leave the 'processing' step, which could contain who-know-what!

3 hours ago, Django said:

the codec isn't helping but it's more than that.. cuz recording to prores externally stills gives out a soft image. the upscaling seems to be the main culprit.

someone in this thread claimed: 

5D3 uses large group pixel bins of 5x5, resulting a 1152x648 RAW bayer, which then gets upscaled to 1080p, that's why it has about 600-700 lines of resolution

 

Interesting - I'll have a read of that thread, thanks.

I am pretty sure that my 700D upscales from 1.7K to 1.9K.  I've directly compared Canon stock 1080 vs ML compressed 1080 at both 1x quality mode and 3x quality mode vs 1.7K RAW upscaled in Resolve 14 and there wasn't enormous differences in terms of detail, just that the detail was smeared in the compressed versions.

The question about the quality of the external feed is whether it is just upscaled or if the bad processing is also applied?  There will be processing applied (colour science for example) so it depends on what else is in the processing apart from that.

2 hours ago, kidzrevil said:

@Django the Canon 5D mark iii is for sure 3x3 pixel binning. I recall reading an article from a canon engineer saying that they didn’t aim to increase the megapixels over the Canon 5D mark ii because with 3x3 pixel binning scales to 1080p without skipping lines. Pixel binning works by combining neighboring pixels and 22.3 with their 3x3 binning equates a 1080p 2mp image.

5760 x 3840 divided by 3x3 = 1920 x 1280 3:2 aspect ratio. The 200 pixels get cropped out for 1920 x 1080 video

What seems to be happening with Canon RAW video in magic lantern is bypassing a lot of the elements that equate the Canon soft resolution. 

Agreed.

Considering that the >22.3MP models aren't magically better I think we can conclude that the upscaling isn't the main issue, and that it's the processing that gets applied.  Combined with the C-line cameras not seeming to suffer these issues, I'd suggest that the poor processing is either a function of the hardware in DSLRs or a deliberate choice in software, perhaps to protect their C-line products.

Regardless, this means that unless they deliberately change this, that 4K seems to be the only likely solution for getting good 1080 out of these cameras (by recording 4K and downscaling to 1080).  Unfortunately this isn't likely to happen soon IMO, considering how slow they've been to introduce it, and considering the awful RS on the M50, which combined with this article from Andrew suggests that it's more of a fundamental problem.

Bummer.  Looks like I'm not only up for a new camera, but also have to change to a new lens system.

[kidzrevil]

@kye what are you shooting with now bro?

[kye]

9 minutes ago, kidzrevil said:

@kye what are you shooting with now bro?

Two setups:

• Canon 700D / ML raw and SD card hack / Sigma 18-35
• Canon XC10

My problem is that the 700D with ML is unreliable, the 700D without ML isn't good enough image quality, the XC10 looks flat, and the AF on both isn't the best.  I was hoping that a Canon DSLR with DPAF might be 'good enough' IQ, but it seems not.

[kidzrevil]

I got rid of the xc10. Too problematic. I went for the mark iii because it hit a lot of bullet points for me and ML works perfectly with it. Canon really cripples their cameras on purpose and the 5D before raw proved that to me @kye

[kye]

36 minutes ago, kidzrevil said:

I got rid of the xc10. Too problematic. I went for the mark iii because it hit a lot of bullet points for me and ML works perfectly with it. Canon really cripples their cameras on purpose and the 5D before raw proved that to me @kye

The XC10 is almost perfect for me, if only it had a faster lens (and the AF to keep up with the resulting shallower DOF).

[newfoundmass]

1 hour ago, kye said:

Awesome, thanks..  So, 4K downscaling + 24Mbps = nice image.  This means that the 50Mbps bit-rate isn't fundamentally flawed.

Yeah, these cameras COULD have very good 1080p even at 24 mbps but the way the video gets processed it never even stood a chance, it's poor whether it's 50 or 500 mbps. 

2 hours ago, kye said:

I'm skeptical that there's a huge difference between H264 encoders (other than the bitrate) and a quick google didn't reveal lots of people comparing encoders so it doesn't appear to be a thing (happy to be proven wrong though!).  

Someone that knows more than me might be able to say otherwise, but I think there is some difference just based on how fine tuned the camera has h264 encoding and the cameras processing power in being able to process it. Less processing power might mean lower quality encoding because it can only handle so much while encoding on the fly. I might be talking completely outta my ass though! 

[kidzrevil]

1 hour ago, kye said:

The XC10 is almost perfect for me, if only it had a faster lens (and the AF to keep up with the resulting shallower DOF).

I would kill for an xc10 with an interchangeable lens. That would be a game changer. The built in lens kinda holds it back...ergonomically its excellent

[Django]

6 hours ago, kye said:

Regardless, this means that unless they deliberately change this, that 4K seems to be the only likely solution for getting good 1080 out of these cameras (by recording 4K and downscaling to 1080).  Unfortunately this isn't likely to happen soon IMO, considering how slow they've been to introduce it, and considering the awful RS on the M50, which combined with this article from Andrew suggests that it's more of a fundamental problem.

Bummer.  Looks like I'm not only up for a new camera, but also have to change to a new lens system.

you might wanna hold up until photokina in september where Canon is expected to announce a high-end (FF) mirrorless. i have a feeling they might get video specs right this time.

[kye]

35 minutes ago, Django said:

you might wanna hold up until photokina in september where Canon is expected to announce a high-end (FF) mirrorless. i have a feeling they might get video specs right this time.

I will be..  partly due to the rumours of impending releases (gold at the end of the rainbow) but also due to there being no good options that seem to suit my (very particular) needs at the moment!!