11 ms·

> I bet the higher resolution makes it easier, with big*ss GPUs to clean up the raw footage. Even if you don't do any other post-processing, downscaled video

by wyager 5y ago

> I bet the higher resolution makes it easier, with big*ss GPUs to clean up the raw footage. Even if you don't do any other post-processing, downscaled video looks better than video captured at sensor resolution. This is because you can implement a steeper low-pass filter (closer to ideal brick-wall filter) in software than you can in actual optics. Hardware lowpass filters tuned to block most stuff above nyquist will also block a lot below it. A 1080p video captured off a good 1080p sensor will have vastly lower amplitude at, say, 500 line pair vertical than a 1080p video captured on a 4k sensor and lowpassed in software with a good filter.

astrange 5y ago
Also, video is not actually bandlimited so the supposedly perfect filter isn’t actually perfect; you can do a better nonlinear one.
- NavinF 5y ago
  What does a better filter look like?
  - astrange 5y ago
    Selective Gaussian blur or hq2x are examples of image filters. Font hinting is an example of a nonlinear downscaling though most people don’t think of it that way. It preserves line sharpness, which is a non-bandlimited feature since lines are infinitely sharp = made of infinite frequencies. Overly sharp linear filters produce artifacts like blockiness and ringing; nonlinear ones would try to avoid that but you could still get more difficult ones like moiré.
    - marcan_42 5y ago
      This is mostly obsolete with High-DPI screens though; we're at the point where such hacks are no longer needed and we can use close to ideal filters (e.g. lanczos, which is a windowed sinc). macOS no longer does font subpixel AA or strong hinting by default for this reason. This is good, because it means we can finally treat computer graphics as properly scalable images without introducing "blurriness". Note that lines were never infinitely sharp; pixels aren't infinitely sharp, they have a shape and the screen construction determines how sharp they look. They are, however, sharper than what you'd get in a nyquist-limited system.
dylan604 5y ago
It was always standard to acquire at higher resolution than being delivered. Hell, they shot 35mm film for SD for decades, 16mm for non-extravagent budgets. The first use of 4k wasn't for 4k delivery, but for single camera shoot of talking head reframing. Full frame for the wide, cropped in for the MED/CU shots. Larger image also allows for cleaner image stabilization. Allows for more detail on pulling keys in post, etc. There are lots of benefits from larger image than delivery.
- kunai 5y ago
  > Hell, they shot 35mm film for SD for decades I'm not sure this is the case. Low-ISO cinema 35mm film (like Kodak VisionT 50D/250D) far exceeds SD resolution. How do you think Star Wars was ever scanned into 4k? Or 2001: A Space Odyssey? Film captures highlight detail logarithmically as well, while digital sensors still haven't really figured out how to render highlights well. This allows greater shadow detail with a skilled cinematographer with no loss of quality in brightly lit scenes. Theoretically, a single 35mm still can be scanned into 4k with no loss in detail, and higher-resolution scans won't lose quality -- there's just dimishing returns and in the end you might just be getting higher-res grain, but it is a far cry from standard definition. Even 16mm can do HD scans just fine. There's a reason Tarantino, Wes Anderson, PTA, Nolan, Sam Raimi, Scorsese, Spielberg, and David Lynch still swear by film (even though many of these directors have dabbled with digital). The idea that film is inherently "low-res" is a really frustrating and common misconception among younger generations unfamiliar with the technology. I find it kind of unbelievable that DPs would shoot 35mm just for an SD scan. That type of use case seems better suited for mini-DV or other contemporary video formats.
  - dylan604 5y ago
    Maybe you misinterpretted what I was meaning. What you quoted from my post was meant as saying they took all of the image from a 35mm frame just to shoot a measly SD image. That's like shooting 4k to deliver SD. It's seemingly way overkill. >That type of use case seems better suited for mini-DV or other contemporary video formats. You realize that people were shooting film for TV production long long before DV was ever invented right?
hammock 5y ago
What does low pass filter mean in the context of video (i understand in audio)?
- astrange 5y ago
  Downscaling or blurring - these are the same thing if you pretend it’s bandlimited. Which it isn’t of course.
  - ryandamm 5y ago
    The signal is typically bandlimited by hardware — the optics and sensor. When it's not, you get weird effects, though that's rare because cameras have optical low pass filters to avoid this. When they don't, like the Canon 5D mark ii when used in video mode (which just threw out most of the sensor's pixels to get to 1080p), you get aliasing producing moire patterns like this: https://youtu.be/dMHDB5fG6ys?t=218 https://youtu.be/dMHDB5fG6ys?t=218
- marcan_42 5y ago
  The same thing: a filter that only lets through low frequencies. A blur effect is a spatial low-pass filter. A motion blur effect is a temporal low-pass filter. What GP is talking about is the antialiasing low-pass filter used when downsampling video to smaller resolutions, or when initially sampling it at the sensor. This is the same exact concept as the antialiasing low-pass filter used when downsampling audio to lower sample rates, or when initially sampling it at the ADC.
  - hammock 5y ago
    Perfect. Thanks!

by wyager 5y ago

astrange 5y ago

Also, video is not actually bandlimited so the supposedly perfect filter isn’t actually perfect; you can do a better nonlinear one.

NavinF 5y ago
What does a better filter look like?
- astrange 5y ago
  Selective Gaussian blur or hq2x are examples of image filters. Font hinting is an example of a nonlinear downscaling though most people don’t think of it that way. It preserves line sharpness, which is a non-bandlimited feature since lines are infinitely sharp = made of infinite frequencies. Overly sharp linear filters produce artifacts like blockiness and ringing; nonlinear ones would try to avoid that but you could still get more difficult ones like moiré.
  - marcan_42 5y ago
    This is mostly obsolete with High-DPI screens though; we're at the point where such hacks are no longer needed and we can use close to ideal filters (e.g. lanczos, which is a windowed sinc). macOS no longer does font subpixel AA or strong hinting by default for this reason. This is good, because it means we can finally treat computer graphics as properly scalable images without introducing "blurriness". Note that lines were never infinitely sharp; pixels aren't infinitely sharp, they have a shape and the screen construction determines how sharp they look. They are, however, sharper than what you'd get in a nyquist-limited system.

dylan604 5y ago

It was always standard to acquire at higher resolution than being delivered. Hell, they shot 35mm film for SD for decades, 16mm for non-extravagent budgets. The first use of 4k wasn't for 4k delivery, but for single camera shoot of talking head reframing. Full frame for the wide, cropped in for the MED/CU shots. Larger image also allows for cleaner image stabilization. Allows for more detail on pulling keys in post, etc. There are lots of benefits from larger image than delivery.

kunai 5y ago
> Hell, they shot 35mm film for SD for decades I'm not sure this is the case. Low-ISO cinema 35mm film (like Kodak VisionT 50D/250D) far exceeds SD resolution. How do you think Star Wars was ever scanned into 4k? Or 2001: A Space Odyssey? Film captures highlight detail logarithmically as well, while digital sensors still haven't really figured out how to render highlights well. This allows greater shadow detail with a skilled cinematographer with no loss of quality in brightly lit scenes. Theoretically, a single 35mm still can be scanned into 4k with no loss in detail, and higher-resolution scans won't lose quality -- there's just dimishing returns and in the end you might just be getting higher-res grain, but it is a far cry from standard definition. Even 16mm can do HD scans just fine. There's a reason Tarantino, Wes Anderson, PTA, Nolan, Sam Raimi, Scorsese, Spielberg, and David Lynch still swear by film (even though many of these directors have dabbled with digital). The idea that film is inherently "low-res" is a really frustrating and common misconception among younger generations unfamiliar with the technology. I find it kind of unbelievable that DPs would shoot 35mm just for an SD scan. That type of use case seems better suited for mini-DV or other contemporary video formats.
- dylan604 5y ago
  Maybe you misinterpretted what I was meaning. What you quoted from my post was meant as saying they took all of the image from a 35mm frame just to shoot a measly SD image. That's like shooting 4k to deliver SD. It's seemingly way overkill. >That type of use case seems better suited for mini-DV or other contemporary video formats. You realize that people were shooting film for TV production long long before DV was ever invented right?

hammock 5y ago

What does low pass filter mean in the context of video (i understand in audio)?

astrange 5y ago
Downscaling or blurring - these are the same thing if you pretend it’s bandlimited. Which it isn’t of course.
- ryandamm 5y ago
  The signal is typically bandlimited by hardware — the optics and sensor. When it's not, you get weird effects, though that's rare because cameras have optical low pass filters to avoid this. When they don't, like the Canon 5D mark ii when used in video mode (which just threw out most of the sensor's pixels to get to 1080p), you get aliasing producing moire patterns like this: https://youtu.be/dMHDB5fG6ys?t=218 https://youtu.be/dMHDB5fG6ys?t=218
marcan_42 5y ago
The same thing: a filter that only lets through low frequencies. A blur effect is a spatial low-pass filter. A motion blur effect is a temporal low-pass filter. What GP is talking about is the antialiasing low-pass filter used when downsampling video to smaller resolutions, or when initially sampling it at the sensor. This is the same exact concept as the antialiasing low-pass filter used when downsampling audio to lower sample rates, or when initially sampling it at the ADC.
- hammock 5y ago
  Perfect. Thanks!