[libcamera-devel] [PATCH 4/5] src: ipa: raspberrypi: Move initial frame drop decision to AGC

[Laurent Pinchart]

Hi David,

On Sat, Dec 05, 2020 at 11:28:49PM +0000, David Plowman wrote:
> On Sat, 5 Dec 2020 at 19:50, Laurent Pinchart wrote:
> > On Fri, Dec 04, 2020 at 04:01:42PM +0000, Naushir Patuck wrote:
> > > On Wed, 2 Dec 2020 at 11:53, David Plowman wrote:
> > >
> > > > Previously the CamHelper was returning the number of frames to drop
> > > > (on account of AGC converging). This wasn't really appropriate, it's
> > > > better for the AGC to do it, which now also knows when exposure and
> > > > gain have been explicitly set and therefore fewer (or no) frame drops
> > > > are necessary at all.
> > > >
> > > > The CamHelper::HideFramesStartup method should now just be returning
> > > > the number of frames to hide because they're bad/invalid in some way,
> > > > not worrying about the AGC. For many sensors, the correct value for
> > > > this is zero.
> > > >
> > > > Signed-off-by: David Plowman <david.plowman at raspberrypi.com>
> > > > ---
> > > >  src/ipa/raspberrypi/cam_helper.cpp  | 6 +++---
> > > >  src/ipa/raspberrypi/raspberrypi.cpp | 8 ++++++++
> > > >  2 files changed, 11 insertions(+), 3 deletions(-)
> > > >
> > > > diff --git a/src/ipa/raspberrypi/cam_helper.cpp
> > > > b/src/ipa/raspberrypi/cam_helper.cpp
> > > > index c8ac3232..6efa0d7f 100644
> > > > --- a/src/ipa/raspberrypi/cam_helper.cpp
> > > > +++ b/src/ipa/raspberrypi/cam_helper.cpp
> > > > @@ -82,10 +82,10 @@ bool CamHelper::SensorEmbeddedDataPresent() const
> > > >  unsigned int CamHelper::HideFramesStartup() const
> > > >  {
> > > >         /*
> > > > -        * By default, hide 6 frames completely at start-up while AGC etc. sort
> > > > -        * themselves out (converge).
> > > > +        * The number of frames when a camera first starts that shouldn't be
> > > > +        * displayed as they are invalid in some way.
> > > >          */
> > > > -       return 6;
> > > > +       return 0;
> > > >  }
> > > >
> > > >  unsigned int CamHelper::HideFramesModeSwitch() const
> > > > diff --git a/src/ipa/raspberrypi/raspberrypi.cpp
> > > > b/src/ipa/raspberrypi/raspberrypi.cpp
> > > > index 0300b8d9..ddabdb31 100644
> > > > --- a/src/ipa/raspberrypi/raspberrypi.cpp
> > > > +++ b/src/ipa/raspberrypi/raspberrypi.cpp
> > > > @@ -192,6 +192,14 @@ int IPARPi::start(const IPAOperationData &ipaConfig,
> > > > IPAOperationData *result)
> > > >         unsigned int dropFrame = 0;
> > > >         if (firstStart_) {
> > > >                 dropFrame = helper_->HideFramesStartup();
> > > > +
> > > > +               /* The AGC algorithm may want us to drop more frames. */
> > > > +               RPiController::AgcAlgorithm *agc = dynamic_cast<RPiController::AgcAlgorithm *>(
> > > > +                       controller_.GetAlgorithm("agc"));
> > > > +               if (agc)
> > > > +                       dropFrame = std::max(dropFrame, agc->GetDropFrames());
> > > > +               LOG(IPARPI, Debug) << "Drop " << dropFrame << " frames on startup";
> > > > +
> > >
> > > All looks good with this change, however, I have a possibly silly
> > > question.  In the previous code, our startup frames would account for
> > > convergence in AGC, AWB, and ALS.  Here we are explicitly accounting for
> > > convergence only in AGC since helper_->HideFramesStartup()  will return 0
> > > by default.  Does it matter? Should each derived CamHelper return a
> > > non-zero number here?
> >
> > Unless I'm mistaken, HideFramesStartup() is meant to report how many
> > incorrectly exposed (and "gained") frames are produced by the sensor at
> > startup, even if exposure time and gain are programmed before starting
> > the sensor. This will certainly impact AWB and ALS as they will have
> > trouble operating if the frame is greatly underexposed, but the sensor
> > is otherwise not involved in AWB and ALS. I thus don't think CamHelper
> > should take AWB and ALS into account.
> >
> > With this new split of responsibilities, with CamHelper reporting the
> > number of frames that are bad (for different reasons, underexposed,
> > incorrect metadata, ...) and the algorithms then deciding how long they
> > need before initially converging, Shouldn't agc->GetDropFrames() be
> > given the HideFramesStartup() value as a parameter, and return a new
> > number, instead of taking the maximum between the two ?
> 
> Yes, I was already planning to change this as it's not quite right,
> but I think I need to change it again. I think the correct behaviour
> is as follows
> 
> Firstly, I'll look into giving AWB and ALSC the same
> GetConvergenceFrames method as the AGC.
> 
> Next, I think the GetConvergenceFrames methods should be given the
> MistrustFramesStartup number. If they actually have no convergence to
> do (e.g. exposure and gain fixed), they return 0. Otherwise they add
> their number to the MistrustFramesStartup number that they were given
> and return that.
> 
> Finally, the dropFrames number is set to the maximum of
> HideFramesStartup and all the GetConvergenceFrames values that we
> obtained.
> 
> This gets the two important use cases right:
> 
> 1. When things aren't fixed, and we give the algorithms as long as
> they say plus the MistrustFramesStartup number, and
> 
> 2. When everything is fixed, all the algorithms should tell us "zero",
> so that we only skip frames if HideFramesStartup tells us to.
> 
> > I also wonder if we then need the hide/mistrust split anymore,
> > especially with the comment in CamHelperOv5647::MistrustFramesStartup()
> > that mentions underexposed frames. Is there still a difference between
> > the two concepts ?
> 
> So I think there is, as discussed above. MistrustFramesModeSwitch,
> though not used in my description above, still has a use too. If you
> were to stop and restart the camera, so the algorithms don't need to
> converge at all, you might still need to tell them to ignore a
> frame's-worth of statistics, if they're thought to be bad in some way
> (otherwise they risk making the algorithm go haywire when it tries to
> process them).

Algorithms consume both statistics and metadata, do I assume correctly
that they would need to ignore both incorrect statistics (caused by
incorrect frames being produced by the sensor) and incorrect metadata
(caused by incorrect metadata produced by the sensor, even if the frame
itself is correct) ? If so, should mistrust frames be the highest of the
number of incorrect frames and incorrect metadata counts, or should it
model the latter only, with the IPA taking the maximum of the hide count
and the mistruct count to decide how many frames to ignore in the
beginning ? I think I have a preference for the latter, as CamHelper
should ideally report raw information about the sensor, without taking
into account how they will be used (same rationale as the one that made
this patch series exist in the first place, moving the convergence time
of the algorithms from CamHelper to the algorithms themselves). What do
you think ?

> > Finally, do we actually need to report a number of frames to drop at
> > startup to the pipeline handler, can't we rely on the algorithms status
> > reported through AeLocked and AwbLocked ? Maybe we should report only
> > the number of frames that are definitely bad based on the CamHelper, and
> > use algorithm status to report initial convergence of the algorithms ?
> >
> > > >                 mistrustCount_ = helper_->MistrustFramesStartup();
> > > >         } else {
> > > >                 dropFrame = helper_->HideFramesModeSwitch();
> 
> It's a good point. In practice we've found that users (OEMs in
> particular) can be very sensitive about getting something on the
> display asap, even if the AE/AWB isn't fully locked. On the other
> hand, it still needs to look "visually locked", avoiding any obviously
> bad frames, so we've always ended up in this sort of compromise
> position where we actually display frames a little before they're
> "technically" locked.

That's a very good point, it makes sense.

> I'll make up another patch set on Monday with both your and Naush's
> feedback, addressed, so thanks for all the suggestions.

-- 
Regards,

Laurent Pinchart