C200: 1080 raw
- Thread starter sebben
- Start date
Cary Knoop
Veteran
Why would you want to do that, if you get 4K RAW you will have a far superior image to start with in post.
Dustin Svehlak
Well-known member
Even if this was possible, it would be pulling a 1080 RAW image from a crop of the sensor or doing line skipping. Either way, it's a major compromise that you would want to avoid.
A better solution for you might be to shoot 4K RAW and then transcode to 1080 or 2K ProRes 444 (I think this may be possible with Canon RAW Development, should definitely be possible with Resolve).
A better solution for you might be to shoot 4K RAW and then transcode to 1080 or 2K ProRes 444 (I think this may be possible with Canon RAW Development, should definitely be possible with Resolve).
Because you would get 4x lower data rate, 1080p which is the most common delivery size and a really good looking image.
Yes, but that is more data to back up and more cards + an extra encoding step which takes time. Surely it is easy enough to do an in camera down scale. 2K supersampled from 4k.
Yes, but that is more data to back up and more cards + an extra encoding step which takes time. Surely it is easy enough to do an in camera down scale. 2K supersampled from 4k.
cpreston
Veteran
There is no such thing as supersampled RAW. The closest thing you are looking for is the RGB 444 12bit that the Canon C300 II records. That is what you want. The 1/4 RAW you are envisioning would be bad.
Recording to an external recorder in 1080p 422 10-bit will provide the quality, latitude, and file sizes you are looking for:
https://youtu.be/jxnWBFne7IE?t=1m27s Start at 1m27s:
You can see he's way over-sharpening the raw, which makes the 422 10-bit external and 420 8-bit internal look soft. While the 4K raw deBayered and downsampled to 1080p in post will indeed carry more detail/information, 10-bit 422 recorded externally will still provide plenty of detail and latitude for flexible color correction and grading.
Remember the C200 isn't a True 4K camera, it's really a 1080p+ camera: 3840*.67 (TRF = True Resolution Factor) = 2572 actual resolution, which isn't much greater than 1920 (to get True 4K we need 3840/.67 = 5.7k Bayer samples. Now you see why the EVA1 and the new C700 FF provide >= 5.7k Bayer sensors).
Those with a C200 and external recorder, you can test this yourself: record 4K raw and 10-bit 422 externally. Bring both clips into your NLE and match sharpness / acutance (in PP CC try Creative/Sharpen). You can do the test both ways: 1080p with 4K downscaled, and 4K with the 1080p upscaled. I've done this test with the 1DX II, which has fairly soft 1080p, as well as with the C300 II, scaling 1080p to 4K. Most can't tell the difference on YouTube. https://www.youtube.com/watch?v=ZIw1CpYKk0c (1DX II 4K and 1080p matched), shot in 1080p and scaled to 4K and sharpened: https://www.youtube.com/watch?v=sor8c7igbFQ (C300 II).
For the GH4 and A7S II, we recorded in 4K 8-bit 420, which when downsampled in post to 1080p looks very good (becomes >8-bit 444. Noise provides a form of dithering that could help with gradients). If you set up your C200 with CLog or CLog3 and colors that are close to the target colors (meaning you're not going to do extreme grading; don't use Canon Cinema Gamut), the internal 4K footage can go a lot farther than you might otherwise think. Ultimately, if outputting for online (8-bit 420) distro, > 8-bit won't help you with gradients (e.g. blue sky, shadows) unless you have sufficient noise/grain to provide some form of dither before truncation to 8-bit. Adding a little noise/grain (PP CC Noise: turn off clipping) before sharpening can help with perceived detail, though for online distro (YouTube etc.) the extra noise can be hard on the encoder resulting in more compression artifacts (for theater / BluRay / non-online-level compression not an issue).
I've done tests with the highly compressed C300 II proxy (8-bit 420 1080p), and was surprised that I could get decent results in post (medium DR tests), even with Canon Log 2 gamma, which is very similar to ARRI LogC and Sony SLog3. Since we had good luck with the A7S II using SLog2, using Canon Log 3 (nearly identical, CLog3 has slightly raised blacks) with 8-bit 420 on the C200 it should also work very well. Also, I've found the best results using Curves and Exposure, Contrast, Highlights etc. (Lumetri) to work much better than using LUTs with all forms of Log and 8-bit footage in PP CC (tested with Canon Official LUTs as well as custom generated (LUTCalc)). You can create a preset that gets you close to the desired look in PP CC, which will be just as fast as a LUT while providing a better end result.
https://youtu.be/jxnWBFne7IE?t=1m27s Start at 1m27s:
You can see he's way over-sharpening the raw, which makes the 422 10-bit external and 420 8-bit internal look soft. While the 4K raw deBayered and downsampled to 1080p in post will indeed carry more detail/information, 10-bit 422 recorded externally will still provide plenty of detail and latitude for flexible color correction and grading.
Remember the C200 isn't a True 4K camera, it's really a 1080p+ camera: 3840*.67 (TRF = True Resolution Factor) = 2572 actual resolution, which isn't much greater than 1920 (to get True 4K we need 3840/.67 = 5.7k Bayer samples. Now you see why the EVA1 and the new C700 FF provide >= 5.7k Bayer sensors).
Those with a C200 and external recorder, you can test this yourself: record 4K raw and 10-bit 422 externally. Bring both clips into your NLE and match sharpness / acutance (in PP CC try Creative/Sharpen). You can do the test both ways: 1080p with 4K downscaled, and 4K with the 1080p upscaled. I've done this test with the 1DX II, which has fairly soft 1080p, as well as with the C300 II, scaling 1080p to 4K. Most can't tell the difference on YouTube. https://www.youtube.com/watch?v=ZIw1CpYKk0c (1DX II 4K and 1080p matched), shot in 1080p and scaled to 4K and sharpened: https://www.youtube.com/watch?v=sor8c7igbFQ (C300 II).
For the GH4 and A7S II, we recorded in 4K 8-bit 420, which when downsampled in post to 1080p looks very good (becomes >8-bit 444. Noise provides a form of dithering that could help with gradients). If you set up your C200 with CLog or CLog3 and colors that are close to the target colors (meaning you're not going to do extreme grading; don't use Canon Cinema Gamut), the internal 4K footage can go a lot farther than you might otherwise think. Ultimately, if outputting for online (8-bit 420) distro, > 8-bit won't help you with gradients (e.g. blue sky, shadows) unless you have sufficient noise/grain to provide some form of dither before truncation to 8-bit. Adding a little noise/grain (PP CC Noise: turn off clipping) before sharpening can help with perceived detail, though for online distro (YouTube etc.) the extra noise can be hard on the encoder resulting in more compression artifacts (for theater / BluRay / non-online-level compression not an issue).
I've done tests with the highly compressed C300 II proxy (8-bit 420 1080p), and was surprised that I could get decent results in post (medium DR tests), even with Canon Log 2 gamma, which is very similar to ARRI LogC and Sony SLog3. Since we had good luck with the A7S II using SLog2, using Canon Log 3 (nearly identical, CLog3 has slightly raised blacks) with 8-bit 420 on the C200 it should also work very well. Also, I've found the best results using Curves and Exposure, Contrast, Highlights etc. (Lumetri) to work much better than using LUTs with all forms of Log and 8-bit footage in PP CC (tested with Canon Official LUTs as well as custom generated (LUTCalc)). You can create a preset that gets you close to the desired look in PP CC, which will be just as fast as a LUT while providing a better end result.
puredrifting
Major Contributor
I was at Canon today and specifically asked about UHD and 1080 CRL. It's not happening, at least for the foreseeable future.
After reviewing more C200 422 10-bit external output: if you need 1080p, record 4K/UHD and downsample in post to 1080p. I haven't seen enough banding reduction (provided by 10-bit, e.g. with sky) to offset the low resolution with both detail (luma) and chroma (color) along with aliasing of the 1080p 10-bit external output. Don't feel too bad- the C300 II aliases in 1080p as well (and 4K on fine detail: again need >= 5.7K to get near alias-free True 4K). Clearly Canon is doing simple pixel skipping type of operations resulting in significant aliasing artifacts. The C200 4K (UHD) 420 8-bit didn't show such issues (see massive aliasing with 1080p on the top edge of the curved orange center section of the walkway):
https://vimeo.com/232042411
Some folks have stated that Canon is simply putting 8-bit in a 10-bit wrapper for external output. Doesn't matter if that's true or not: the low resolution and aliasing would be a full stop for me. For the light house shot, C100 matched pretty closely to the C200 4K internal downsampled to 1080 (C200 has slight edge). For the same shot, when viewed at 1080p (video uploaded as 4K), the 12-bit 444 raw material was more detailed as expected, however for any content beyond a test shot, no one would likely notice a difference. When viewed in 4K, the difference it significant (some might say too sharp: he used CRD Sharpen 5).
The C100 and C300 both provide superior 1080p to the C200 and C300 II: they use 4K sensors downsampled to 1080p (R & B taken directly from the Bayer array (not actually downsampled: direct sampled and not Bayer processed, so no possibility of Bayer-related color aliasing), G is downsampled with a simple box (averaging) filter of two sensels (to become one pixel). No actual deBayering takes place, so no Bayer artifacts!). Canon is clearly not using that algorithm with the C200 and C300 II for 1080p. The guess as to why would be marketing: folks would question why 4K with only a 4K sensor (2.5K after deBayer) doesn't look much better vs. 'at the Nyquist limit' True 1080p (from 4K, not deBayered) if using the same algorithm used with the C100/C300. So they provided a low-quality 1080p so 4K looks visibly better. The C200/C300II are still better all around than C100/C300 (noise, DR, AF etc.), however I won't purchase another (marketed as) "4K" camera with less than a 5.7K sensor (and probably not less than a FF sensor).
Many of us shot for years on the GH4 and A7S II in 4K 100Mbps downsampled to 1080p without any major banding issues. For some folks Panasonic VLog created issues in the same way Sony SLog3 created issues: they're just a bit too flat for 8-bit in many cases (VLog and SLog3 are nearly identical curves). However, I never had any major issues, and don't recall reading about issues from others, with Sony SLog2 with the A7S II. And guess what SLog2 is a nearly identical too? Canon Log 3. So you should be fine with 8-bit on the C200 with CLog3 shot internally. If you find banding issues and/or extreme grades cause issues, you have the option to shoot raw for those cases. Add a little Gaussian blur to the downsampled 4K raw to 1080p to match the internal material and it should match fine.
https://vimeo.com/232042411
Some folks have stated that Canon is simply putting 8-bit in a 10-bit wrapper for external output. Doesn't matter if that's true or not: the low resolution and aliasing would be a full stop for me. For the light house shot, C100 matched pretty closely to the C200 4K internal downsampled to 1080 (C200 has slight edge). For the same shot, when viewed at 1080p (video uploaded as 4K), the 12-bit 444 raw material was more detailed as expected, however for any content beyond a test shot, no one would likely notice a difference. When viewed in 4K, the difference it significant (some might say too sharp: he used CRD Sharpen 5).
The C100 and C300 both provide superior 1080p to the C200 and C300 II: they use 4K sensors downsampled to 1080p (R & B taken directly from the Bayer array (not actually downsampled: direct sampled and not Bayer processed, so no possibility of Bayer-related color aliasing), G is downsampled with a simple box (averaging) filter of two sensels (to become one pixel). No actual deBayering takes place, so no Bayer artifacts!). Canon is clearly not using that algorithm with the C200 and C300 II for 1080p. The guess as to why would be marketing: folks would question why 4K with only a 4K sensor (2.5K after deBayer) doesn't look much better vs. 'at the Nyquist limit' True 1080p (from 4K, not deBayered) if using the same algorithm used with the C100/C300. So they provided a low-quality 1080p so 4K looks visibly better. The C200/C300II are still better all around than C100/C300 (noise, DR, AF etc.), however I won't purchase another (marketed as) "4K" camera with less than a 5.7K sensor (and probably not less than a FF sensor).
Many of us shot for years on the GH4 and A7S II in 4K 100Mbps downsampled to 1080p without any major banding issues. For some folks Panasonic VLog created issues in the same way Sony SLog3 created issues: they're just a bit too flat for 8-bit in many cases (VLog and SLog3 are nearly identical curves). However, I never had any major issues, and don't recall reading about issues from others, with Sony SLog2 with the A7S II. And guess what SLog2 is a nearly identical too? Canon Log 3. So you should be fine with 8-bit on the C200 with CLog3 shot internally. If you find banding issues and/or extreme grades cause issues, you have the option to shoot raw for those cases. Add a little Gaussian blur to the downsampled 4K raw to 1080p to match the internal material and it should match fine.
AndreeOnline
Veteran
Way back, when I was doing some military training in northern Sweden during a navigation exercise with map and compass, one guy found that what he was seeing with his own two eyes didn't seem to fit with the map he was using. In an effort to correct the misalignment he got down on this hands and knees trying to turn the whole planet around to lineup with the map.
I think he's still out there.
Must have been around 24 years ago...
Truth always finds a way.
AndreeOnline
Veteran
I think another level of compression would be better. Something like 8-10x.
I remember reading this whitepaper around the time the C300 II was released: http://www.thedithouse.com/on-set-d...-dailies-systems/canon-digital-slr-workflows/ . They state a lot of interesting things related to sampling theory (including a claim that the sensor data is upsampled before processing (this is the 4K data! Upsampled to what resolution?)). Then the data is processed into RGB then YUV (YCrCb) etc. with statements of increased detail, reduced aliasing, and reduced noise. The noise reduction is certainly working, as even at -1 NR, the blacks are very impressively clean (C300 II). However aliasing and detail performance in 1080p isn't matching the whitepaper (C300 II and C200). If the C200 is using the same sensor and algorithms, how is it possible that the C100 is producing better 1080p than the C200? (including 10-bit 422 external vs. C100 8-bit 420 internal)
Again, for those looking for the highest quality 1080p from the C200, you'll get the best results shooting 4K and downsampling to 1080p in post. This worked great with the GH4 and A7S II and will work great for the C200, with relatively small files, which will be effectively 1080p 444 >8-bit (Y = 10-bit, U,V 8 each for 26/3 = ~8.67bits (plus minor losses converting to/from RGB). With sufficient scene noise (or added in post before downsampling) the extra bits can help with banding)). Remember extra bits e.g. 8 vs. 10 aren't going to help DR, only help reduce quantization issues with de-logging and extreme grading before final quantization to 8-bit 420 for delivery (unless going to cinema, PC playback with a fully 10-bit display system (rare), or for the latest HDR systems (including YouTube)). For certain fast motion shots, external recording to an ALL-I codec can be helpful (though with motion blur and not pixel peeping still frames, codec breakup may not really be an issue if not too severe).
Again, for those looking for the highest quality 1080p from the C200, you'll get the best results shooting 4K and downsampling to 1080p in post. This worked great with the GH4 and A7S II and will work great for the C200, with relatively small files, which will be effectively 1080p 444 >8-bit (Y = 10-bit, U,V 8 each for 26/3 = ~8.67bits (plus minor losses converting to/from RGB). With sufficient scene noise (or added in post before downsampling) the extra bits can help with banding)). Remember extra bits e.g. 8 vs. 10 aren't going to help DR, only help reduce quantization issues with de-logging and extreme grading before final quantization to 8-bit 420 for delivery (unless going to cinema, PC playback with a fully 10-bit display system (rare), or for the latest HDR systems (including YouTube)). For certain fast motion shots, external recording to an ALL-I codec can be helpful (though with motion blur and not pixel peeping still frames, codec breakup may not really be an issue if not too severe).
AndreeOnline
Veteran
Don't they just mean the process where each color component of the Bayer pattern is extracted (R, G1, G2, B), the "holes" are filled in by averaging (which leads to the benefits you mentioned—especially aliasing), before the individual components (now without holes in them) are reassembelled?
I've always found this method to make sense and seem effective in terms of "working with what you've got".
Sure the deBayer process interpolates pixels, which is upsampling, however that's just part of the standard algorithm. The way they wrote and diagrammed, it appears deBayer is a separate process after upsampling (it appears they are deBayering this time (vs. C300 I) re: apparent chroma aliasing with C200 footage):
Whatever they're actually doing, while it works great for low noise, it's got both luma and chroma aliasing issues (10-bit 422 1080 C200), and it's soft/low-detail! Take a DSLR still (5+K sensor) and drop it into a 1080p or 4K sequence/timeline, then scale down to frame size. Compare to the same shot with your video camera. The DSLR still downsampled to 4K (or 1080p) shows what True 4K / HD should look like- you might be surprised by what you see.
Whatever they're actually doing, while it works great for low noise, it's got both luma and chroma aliasing issues (10-bit 422 1080 C200), and it's soft/low-detail! Take a DSLR still (5+K sensor) and drop it into a 1080p or 4K sequence/timeline, then scale down to frame size. Compare to the same shot with your video camera. The DSLR still downsampled to 4K (or 1080p) shows what True 4K / HD should look like- you might be surprised by what you see.
AndreeOnline
Veteran
Well, a normal deBayer would interpolate between the various color components, for instance between a red, two greens and a blue, but each underlying matrix would only be quarter resolution, if we count both greens as separate.
I'm interpreting Canon's white papers as they are filling in each of the four color matrices via interpolation so that you get four full resolution color components. And then they are combined together again. I would consider this extra step the upsampling. Standard deBayer would work directly on the RGGB grid.
I think the image from the white paper is pretty clear, actually.
indiawilds
Veteran
Wow! Had a good laugh. If/when ever we meet, I will buy you a beer.
Linearly interpolating the Bayer grids followed by a median filter ("False Color Suppression") is just about the fastest / lowest compute cost method of... deBayering (demosaicing)! "Raw" typically means raw Bayer data. In their chart Raw means post deBayer (RGB before color gamut and gamma processing). The entire "Over Sampling" block is just the standard (fast) way of deBayering- nothing new: why the implication of something novel? Pre-filtering is some kind of low-pass filter (blur) and downsampling could by simple decimation (just grab 1920x1080 pixels) or maybe it's just simple Box filtering (averaging). The results don't look like high quality downscalers such as bicubic or Lanczos etc. are used (which we can compute on our cell phone hardware!).
The puzzle here, is why does the UHD 8-bit 420 look pretty good (slightly better than the C100 HD from 4K) and the 2K 8-bit 420 and 10-bit 422 look nearly the same: low resolution with both zippering (luma aliasing) and color artifacts (chroma aliasing)?
C200 2K 8-bit 420:
C200 External 10-bit 422:
C100 8-bit 420 (low bitrate (24Mbps) AVCHD) and C200 UHD 8-bit 420:
There is an improvement in detail from the C100 1080p to C200 UHD. What is the reason for making the C200 HD look so much worse than the C100 HD? Related to the OLPF being tune for 4K instead of 1080p? However, why not just downsample the 4K image to 1080p- it's a very fast/simple operation?
C200 2K 8-bit 420:
C200 External 10-bit 422:
C200 4K Raw:
The 4K raw shows the quality boost of a much more advanced deBayer algorithm. When working with 5D3 raw, which needed as much help as possible (1080p Bayer grid), I found AMaZE looked pretty good compared to whatever method Adobe uses in Adobe Camera Raw (ACR): https://rawpedia.rawtherapee.com/Demosaicing . For fast hardware implementations, there are a few more variants: http://www.fastcompression.com/products/debayer/debayer.htm
In summary, the C200 UHD looks OK (no major artifacts but limited in real detail with a 4K sensor targeting 4K output (not True 4K/UHD)). If targeting 1080p, shooting in UHD and downsampling will provide superior results vs. shooting 2K/HD internal or external. While I've seen aliasing on the C300 II (both HD and UHD, including zippering), it hasn't been very significant, and have not seen artifacts such with the C200 and light house. The C200 has more computing power than the C300 II, so it's not clear why this is happening, beyond perhaps marketing so UHD looks significantly better than 2K/HD in the same camera?
By increasing acutance, we can make the C100 HD appear sharper than the C200 UHD (on a quick glance):
However looking closely, it's not real detail, and the C200 is a much cleaner image and looks more organic/filmic. This is why we want .67 factor (or better) oversampled Bayer sensors! (at least 5.7K for UHD)
However looking closely, it's not real detail, and the C200 is a much cleaner image and looks more organic/filmic. This is why we want .67 factor (or better) oversampled Bayer sensors! (at least 5.7K for UHD)
Here's 1080p on the C300 II. I've never compared directly to the C300 I, however the C300 II looks like soft 1080p (4K is soft due to not having a 5.7K+ sensor). In this video I didn't bother to blur the background plate, so it looks like the camera guy focused on the pyramids or shot at F32 lol. The background plate represents what full 1080p should look like, detail wise (captured from a much higher resolution sensor than 1080p, then downsampled):
https://www.youtube.com/watch?v=FDY4FEFQ1o0
I sharpened as much as possible, before it started to look bad, and it's still pretty soft. However, there are aliasing artifacts around the glasses and the UFO (high contrast) edges are starting to look bad/brittle (over sharpened). This example is really about aliasing, as part of the softness comes from YouTube's 1080p compression. If we take a C300 II 1080p video shot in 12-bit RGB 444 (RGB 444 is the most important element, not 12-bit), scale it to 4K, then post sharpen and deliver in 8-bit 420 4K/UHD, YouTube does a much better job:
https://www.youtube.com/watch?v=sor8c7igbFQ
RGB 444 provides all the color necessary for native 4K/UHD 420, so only luma is sub-sampled, and that's why it can look pretty good scaled to 4K and sharpened.
This example is perhaps a bit oversharpened (specular glints looks digital/crunchy (need a smart sharpen in PP CC that leaves high-contrast edges alone- if full (same APIs available as native plugins) GPU accelerated plugins were possible by third parties I'd consider writing it myself (their third party SDK is a mess (need working examples for both CUDA and OpenCL, perhaps Metal)))).
For C200 users, I'd expect 4K/UHD 8-bit 420 internal to look pretty similar to 4K 12-bit raw where both are downsampled in post to 1080p. I haven't used CRD- does it provide any NR (doesn't look like it from screenshots)? If not, then internal 4K has another advantage: pretty excellent in-camera NR (can save a lot of time in post using Neat Video etc.).
https://www.youtube.com/watch?v=FDY4FEFQ1o0
I sharpened as much as possible, before it started to look bad, and it's still pretty soft. However, there are aliasing artifacts around the glasses and the UFO (high contrast) edges are starting to look bad/brittle (over sharpened). This example is really about aliasing, as part of the softness comes from YouTube's 1080p compression. If we take a C300 II 1080p video shot in 12-bit RGB 444 (RGB 444 is the most important element, not 12-bit), scale it to 4K, then post sharpen and deliver in 8-bit 420 4K/UHD, YouTube does a much better job:
https://www.youtube.com/watch?v=sor8c7igbFQ
RGB 444 provides all the color necessary for native 4K/UHD 420, so only luma is sub-sampled, and that's why it can look pretty good scaled to 4K and sharpened.
This example is perhaps a bit oversharpened (specular glints looks digital/crunchy (need a smart sharpen in PP CC that leaves high-contrast edges alone- if full (same APIs available as native plugins) GPU accelerated plugins were possible by third parties I'd consider writing it myself (their third party SDK is a mess (need working examples for both CUDA and OpenCL, perhaps Metal)))).
For C200 users, I'd expect 4K/UHD 8-bit 420 internal to look pretty similar to 4K 12-bit raw where both are downsampled in post to 1080p. I haven't used CRD- does it provide any NR (doesn't look like it from screenshots)? If not, then internal 4K has another advantage: pretty excellent in-camera NR (can save a lot of time in post using Neat Video etc.).