[libcamera-devel] softISP for libcamera

[Laurent Pinchart]

On Wed, Dec 06, 2023 at 04:45:21PM +0100, Pavel Machek wrote:
> Hi!
> 
> > > > > That will result in shooting at 1/15, ISO50 -- resulting in blurry
> > > > > photos (motion) -- even with sensors that could shoot at ISO500
> > > > > without getting grainy. RPI pipeline has support for this, with config
> > > > > file saying "stay at ISO50 to 1/60, then move iso up to ISO200 @ 1/60,
> > > > > then start adjusting exposure again". It would be nice to reuse it.
> > > > 
> > > > Let's please not refer to "ISO" in technical discussions. The ISO 12232
> > > > standard is pretty much useless, "ISO" values are only meant for
> > > > marketing.
> > > 
> > > Well, I believe "marketing ISO" is good enough to demonstrate my point
> > > here :-). It is familiar "unit" to people doing photography. What
> > > would you suggest as replacement?
> > 
> > Just analog gain. There's nothing anywhere that maps a x1.0 analog gain
> > to ISO 50, 100, 200 or any other value.
> 
> Exactly, and that means that ISO still has its place, right?
> 
> 1/125s, ISO100, f/3.5 corresponds to known lightning conditions across
> range of cameras (including film ones).

ISO 100, as a sensitivity value for digital cameras, doesn't mean
anything. It's an arbitrary value picked by the manufacturer. For film
cameras it could be used to compare sensitivity of films, for digital
cameras, it's meaningless.

> 1/125s, f/3.5, a.gain 1x does not correspond to anything, unless you
> know camera details.
> 
> > > Anyway "minimal possible gain" is too simple solution for many cases.
> > 
> > That I don't dispute :-)
> 
> Ok.
> 
> > Also note that there's no reason for going lower than x1.0, even if the
> > sensor supports lower values.
> 
> I'm not sure I agree here. x0.5 gain will still have a tiny bit lower
> noise than x1.0, no? (Plus there's place for it if you have too much
> light and can't make shorter exposure).

You're not going to meaningfully decrease your SNR, and you won't avoid
saturation by lowering the analog gain below x1.0.

-- 
Regards,

Laurent Pinchart