Originally Posted by ddgdl
I asked this in another thread but figured this is probably the better one to ask it in (though you may be biased!)---
I know madshi is working on real time dynamic mapping by looking ahead a few frames, which markmon and others have said will obviate the need for measurements. But won't pre-measuring the entire file always give the best results, at the very least by offloading some of the real-time processing work that would have to be done otherwise? Am I missing something?
Well honestly, without being biased
, it's pretty obvious that *in theory" you should be able to have much better results than ANY version of the LIVE algo IF you know all the frames in advance.
Now, having said that, it of course depends on how good in the LIVE algo and how good is the "Measurement file based" algo.
But, if both are given the same developement effort, "measurement file based" will always beat by quite a margin the LIVE algo.
So for the *maybe* biased part:
the current philosophy of the "Dynamic target" live algo is to use the "FALL algo" and then reset the target nits for each new detected scene.
This means that if 2 people are talking to each other in the same room, and the camera keeps switching every 2s between the two, the target nits will get reseted at each camera change. IF the FALL is very different between the 2 camera perspective, then you will observe large target nits change at each camera change.
--> This might be disturbing since while in the same "conversation" the brightness will change up and down
Ideally, you would like for such "conversation between 2 people" to be handled as one single scene, so that you don't have strange brightness jump between the same "chapter".
--> This is where the "measurement file" has a huge advantage: you can basically detect such situation by looking ahead.
-->In our tool working with the measuement file, we "try" to only reset the target nits only large "chapter" change (large FALL change). "Chapter merge" is here for that.
We also made an improvement in the V3.7.2 (not released to you guys yet
), where we basically handle "sandwich scene" to help with this kind of situation where the camera switch back and forth, but you actually want to keep the same target so that the brightness does not change too much.
Also while in a "SCENE", the live algo is able to change the target nits based on a predefined speed scaled to the current delta to the desired frame peak. But Live algo does not known the future.
With the measurement file, we currently use a centered rolling avg so that we can better prepare for the future with a smoth transition to the future required target nits.
So while I am very happy that the LIVE algo now included the dynamic target nits, and while the LIVE algo may please most people and avoid the 'hassle" of generating measurement file, the "measurement file" will still be beneficial to be "smarter" about the "when" do you reset the target nits so that the relative change between "friend scenes" is still respected.
Also, we should not forget that Madshi first purpose of the measurement file was to know in advance the frame peak so that you would never tone map to a peak smaller than the current frame peak which would result in indesired clipping. The live algo may lead to uncontrolled clipping because of the "fixed" moving speed. With the measurement file, it's always "perfect", because it can react in advance (it knows the future) and raise the tone mapping peak target progressively so that when it does reach the frame with the highest peak, it still does not clip.
And finally, you can probably do much more than what we are doing at the moment with the measurement file to make it even better.
Edit: and yes you are right at the very least you can unload your gpu and save some precious ms if you use the already pre-analysed and post-processed measurement file.