C200: 1080 raw
- Thread starter sebben
- Start date
Go to 57s. The first part is cropped in slightly: chose that video to show aliasing more clearly around the glasses / UFO. Also YouTube 1080p is very compressed (for anything where you want to show high detail, upload to YT as 4K).
11 years ago, this was the best YT could do with fast motion (30fps):
https://www.youtube.com/watch?v=zOJh7wpCxXM
9 years ago (filmed with a Canon HF11 pointed at the screen lol):
https://www.youtube.com/watch?v=QxF3jg05UqM
Now in 2018, we can to 4K60p with pretty amazing quality compared to prior years! (looks like Chrome is needed to see 4K60p (MacOS))
https://www.youtube.com/watch?v=oYg6hIo3vJA
Might be a fun creative project to do a project with a 4K60p timeline/sequence, record most everything in 4K24p (just drop into the 60p timeline (frame sampled, will look fine)), then for effect/story, add in 60p elements, not slowed down (even upsampled+post sharpened 1080p to 4K). 120fps slowed to 60p would be velvet smooth. Kinda similar to using different aspect ratios for effect in the same film/video to help with story/emotion.
In high quality debayer algorithms the underlying resolution applied to all color channels is full luminance resolution of the sensor. High quality debayer alogrithms don't produce color fringing or aliasing, but the built in debayer processing in video cameras is compromised for speed over quality.
Nice organic 1080p images: Canon DSLR + ML RAW? Last image shows edge artifacts in high-contrast areas (looks too sharp / discontinuous). If you take a 5D3 still (5.7K, even a JPG) downsampled to 1080p and compare to an ML RAW 1080p frame with the best available deBayer method, the difference will be pretty massive. While the 1080p image will look nice and organic (as in your examples), the 5K downsampled to 1080p image will also look nice and organic (even more so), and will also contain an incredible amount of real detail and zero over-sharpened edges or other aliasing artifacts.
Regarding higher quality deBayer in camera- current general-purpose hardware technology (GPUs, AI processors) available in cell phones have more than enough compute power to perform top-of-the-line deBayer in camera. These are general purpose chips, which can be updated just as easily via firmware as with FPGAs (used in ARRI cameras). FPGAs while powerful are also large and power hungry (and comparatively expensive). For large companies like Canon, Sony, Panasonic, etc., they can have ASICs created which aren't general purpose, and are not programmable. However, they are low cost at production scale, and can be low power. Canon's DIGIC chips are ASICs, so clearly Canon can make them to do anything they wish. ASICs are being used for AI projects, so it's clear ASICs can be competitive with GPUs/AI-processors for image processing. The question is why isn't Canon doing higher quality deBayer (and other processing)? Haven't seen any 4K UHD C200 footage with significant artifacts yet, so perhaps they're doing a pretty good job with 4K. Then the question becomes why does 10-bit 422 1080p have major artifacts?
5D3 ML 1080p raw, processed with Adobe ACR. Gorgeous color (much easier to achieve great looking color* than the 1DX II, C200, C300 II), organic, though a bit soft compared to True HD.
https://vimeo.com/92437795
* (something changed in Canon's color science after the 5D3)
Those three frames are from my Digital Bolex @ 24fps, 180 degree shutter, ISO200.
Top to bottom:
1. Cosmicar 12.5mm C-mount prime, f/4. Processed in Resolve.
2. Minolta 35-70 f/3.5 macro zoom in macro mode, f/5.6 if i remember correctly. Also processed in Resolve.
3. Angenieux 12-120 HEC zoom with +1 diopter macro lens attached. f/4. Processed in Raw Therapee. Yes a bit oversharpened at the edges.
The Bolex has a well matched OLPF, so it does read a bit soft but moire and aliasing artifacts are rare to see. The upside to that is you can add quite a bit of sharpening, especially fine radius micro contrast, to bring out detail without it getting that hard edged video look. I like to push sharpening right to the edge of visible aliasing on edges then add just a touch of low radius Gaussian blur to smooth high contrast edges while retaining the detail.
Downscaling from higher resolutions does help for sure.
Raw Therapee has a selection of much better DeBayer algorithms than Resolve but the processing power they require goes up exponentially. Resolve will run raw timelines full resolution at 24fps on my i7 quad core. That single frame processed in Raw Therapee took about 4 seconds.
Last edited:
Cool- the Bolex looks very organic with nice skin tones- love the CCD look. Oversharpened edges: would be a cool feature to have- selective blur; only affects high contrast (high frequency) / sharp edges. Would not be a complex filter: edge detection to create a mask, then apply Gaussian/Box blur to the mask in amount proportional to the mask intensity.
These guys have a demo for fast GPU deBayer (high-quality methods): http://www.fastcompression.com/download/download.htm . http://www.fastcompression.com/products/debayer/debayer.htm . At some point NLEs should get more advanced deBayering, perhaps reserving the highest quality when final rendering (still GPU accelerated). Downsampling a higher resolution sensor puts less 'load' on the need for ultra-high quality deBayering.
AndreeOnline
Veteran
Yes, that's essentially all I'm trying to say: that Canon is making a point of applying a method closer to bicubic than bilinear interpolation. I'm not sure of the exact method they are using and I also don't know what is considered to be the "standard" method, if there is such a thing, among other video cameras.
And then you have even more exotic and advanced methods like in Raw Thereapee that can let the full power of a workstation churn on a single image for a bit to squeeze everything out of it. I did experiment with Raw Thereapee a few years back on my photos which led to a local peak in knowledge about various advanced algorithms, but most of that has faded over the years I'm afraid.
Even if they did excellent Bayer grid interpolation (such as Lanczos), it's not going to provide much of an improvement. The higher end deBayer methods use various forms of statistics (forms of correlation) to compute fairly accurate luminance for every pixel, while at the same time preventing aliasing (both luma and chroma). Here's one example: http://www.academia.edu/2666728/Bayer_Pattern_Demosaicking_Using_Local-Correlation_Approach . See page 32, where some methods have obvious luma and chroma aliasing issue, and their method (l) looks perfect. These statistical correlation methods are indeed far more expensive to compute on a general purpose CPU, however GPUs can do it fairly quickly, and a custom ASIC could be very fast.
Whatever Canon is doing for 4K/UHD, looks pretty good (C200 lighthouse example). It's unclear what they are doing for 2K/HD 8-bit 420 or 10-bit 422 (both have aliasing issues). The C300 II has aliasing issues in 4K (the C200 will as well with high-frequency scene detail (4K Bayer grid will alias when outputting 4K (Bayer 4K -> RGB 4K is == Nyquist frequency: we need > Nyquist to prevent aliasing (along with appropriate OLPF and post filtering)). We found .67 to be a good predictor for real-world performance after some long forum debates (my initial estimate with .5 was based in how good the Sony F65 looks ("8K" to get 4K), revised to .67 after further analysis and multiple forum discussions). For HD, we need (at least) 1920/.67 = 2.87K. ARRI 2.88K for HD should now make sense (private discussion with ARRI shows they are well aware of the .67 factor, so we reached the right number)). Same for 4K 3840/.67 = 5.7K (minimum, more is better)). So the new Canon FF with 5.9K sensor for 4K should look stunning (sounds like it does: puredrifting was impressed after testing it (review coming)). Finally a True 4K camera from Canon!
