Hacker News Reader

Comment by endorphone | Hacker News Reader

They again are sticking with EIS, and again the demos are of extremely well lit scenes.

EIS solutions don't hold up quite as well in lower light, where OIS absolutely shines. Video, just like photos, has individual frames that have exposure times, and in lower light the longer exposure time == garbage source frames that no amount of EIS can fix. But expect a tonne of extremely well lit examples pretending that it's a cure all.

The best solutions marry EIS and OIS. That's exactly what the iPhone does.

View on HN

The respective pros and cons of EIS & OIS don't change with light level. EIS works by cropping part of the image which results in a narrower FoV and a lower resolution, however these effects occur regardless of how bright the scene is.

The additional moving parts of OIS systems make them less suitable for ruggedized cameras.

Video is a series of photos. Captured by the photo sensor identically to taking photos, but for bandwidth and time reasons usually downsampled and with a max exposure time of the framerate.

In low light that 30 FPS camera will increase the exposure time to up to 1/30th of a second. In a bouncing, rapidly changing camera this yields a smear of a photo. OIS combats this by limiting capture motion during most movements.

There is absolutely an enormous difference between the two, and it yields the classic garbage in = garbage out problem.

You're ignoring the effects of motion blur. Consider a smoothly panning camera in a poorly lit scene. OIS can move the lens to cancel out the pan so you don't get blurring.

EIS can only do it if you take multiple photos at a much higher frame rate (I think the Pixel 2 does this but it is very uncommon, and probably too power intensive to do for 4K video).

You're ignoring the application.

OIS systems move a subgroup of lens elements to compensate for movement of the camera. The degree of that movement is very limited. The main purpose of OIS is to approximate a stationary camera platform from a near-stationary one. Many OIS systems get turned off for panning - while more sophisticated ones support being turned off in the panning axis - specifically because they don't handle movement outside of a very limited range. For applications where the camera is intended to be in-motion the appropriate mechanical solution is using a gimbal, not OIS.

The main purpose of gimbals or EIS is not to produce sharper individual frames but to reduce jarring across a sequence of frames by smoothing the camera movement - gimbals do it physically while EIS simulates it after the fact by using a subset of the total sensor area.

At 30fps the longest possible per-frame exposure is 1/30s. An action camera-wearing athlete moving at a leisurely 20mph will travel about a foot in 1/30s - and that's just movement in one axis. OIS is not going to help with per-frame motion blur when the camera moves that much or more during the exposure.

OIS is not going to help with per-frame motion blur when the camera moves that much or more during the exposure.

A camera rotating or tilting just a few degrees has magnitudes more of an impact than the fastest possible athlete. And yes, of course OIS "turns off" during panning (and the resulting image blur is not only consequent, it is expected and natural), as it would counteract the effect. Just to be clear, I spent two years building a commercial image stabilization solution -- I'm pretty educated on what I'm talking about.

The value of OIS + EIS is that with OIS the vast majority of frames have the stablest possible capture, while EIS can effectively fix what OIS misguessed (e.g. the beginning of a pan that OIS tries to counteract). This is exactly what the iPhone does.

> A camera rotating or tilting

> I'm pretty educated on what I'm talking about.

Then you should know that tilting is rotation. Unless by rotation you mean specifically around the optical axis to differentiate from pitch and yaw... in which case you should know that such rotation is uncorrectable by optical systems - though it is correctable by sensor-shift systems.

I think the point you're trying to make with "a few degrees has magnitudes more of an impact" is that rotation has more impact than shift. I agree that the rotation caused by moving the lens a few millimeters has more impact than shifting the camera a few millimeters for sufficiently distant subjects. This is why early IS systems were focused on mitigating pitch and yaw as opposed to shifts: moving lens groups to counter millimeters of rotation is within the physical capabilities of a system that can fit within a lens and it gives a lot of bang for the buck in a narrow sweet spot. Rotational movement is most prevalent in hand-held cameras while operating the camera - for example pressing the shutter button - and less so when it's just being held. This is why there was a standard trick in still photography of using bursts of continuous shooting to get clearer captures: the first frame is impacted by sudden pitch but subsequent ones are much less so. OIS does an excellent job of addressing that first frame blur.

Due to inertia, when a camera is attached to a larger object in motion (athlete, surfboard, vehicle, etc.) random rotational movement is much less pronounced while continuous shifting is. Correcting sustained directional movement - be it rotational or shift - is not something OIS systems are designed to handle. They are meant to correct random movement around a mean, essentially they provide an instant mechanical "reversion to the mean" of sorts. Correcting tens of centimeters of shift is well outside their capabilities and shifting a camera tens of centimeters during exposure will absolutely result in noticeable blur. It doesn't remotely require "the fastest possible athlete" to achieve that level of shift during a 1/30s exposure.

The point is not that EIS is better than OIS - the systems serve different purposes. The point is that OIS doesn't bring as much value to action camera use cases and it comes with the significant downside of reduced durability.

5-axis OIS system absolutely correct for rotation (e.g. roll), as of course sensor-shift and rotation is a component of OIS.

The point is not that EIS is better than OIS - the systems serve different purposes

They serve remarkably similar purposes and can be complimentary. It's actually interesting because your claims sound like what Google said with the Pixel (1) and its completely substandard EIS-based solution. With the pixel 2 they added OIS and complement it with EIS. Everyone eventually gets to OIS + EIS. But while they're trying to be cheap we hear the spiel about why they just don't need OIS.

Because EIS alone is substandard unless you have a fantastically fast sensor or perfect lighting. Which is why EIS samples are always on the brightest possible days. It isn't by accident.

EIS is more akin to a poor-man's gimbal than it is to OIS, and that's the purpose it serves in GoPros. One could argue that for sufficiently small range of movement a gimbal and OIS can serve similar purposes for stills photography - a gimbal after all is a form of external stabilization. However EIS is not seriously applicable to stills photography as it does nothing for a standalone capture and OIS cannot replace a gimbal (or EIS) for video capture other than in very narrow scenarios where static support would generally be a better solution.

It's obvious that OIS + EIS > EIS. However it's not that case that OIS > EIS for all applications. One could argue that given just OIS one could apply EIS in post, but that's just a different implementation of OIS + EIS.

As far as I know, the GoPro is fixed-focus - there's no moving parts in its optical system. Adding mechanical stabilization whether optical or sensor-based would reduce the durability of a system of which one of the main criteria is being able to withstand repeated impacts. Maybe it's possible to build a sufficiently ruggedized OIS solution, but it's not clear whether the expense or size/weight of that would make it viable for action cameras.

Reply on news.ycombinator.com