Feature Request: Adjustable Sample Size

[zhdenny]

I understand 10 samples spread across the timestamps is probably just fine....but it would be nice if the sample size was user-adjustable via a command-line flag.

Thanks

[cryptochrome]

Hey there,

thanks for the suggestion. I actually have this on my internal roadmap already, just not sure about the specifics yet. What do you think would be a good sample size we should use by default?

I will make it adjustable in a future version.

[zhdenny]

I don't know what a good default would be. I figure you came up with 10 for a reason so maybe just keep that as default.

[cryptochrome]

No worries. I had only 3 samples before, and that was way too little, as it turned out. So increased it to 10, which not delivers reasonable results. I will make this user adjustable up to 30 (more than that will have diminishing return). Added to public roadmap now.

Until this is implemented, you can use -inspect (which will do a full frame by frame analysis).

[cryptochrome]

An update on this:

I am putting this a bit on the backburner, as I am currently exploring other options to increase the accuracy of the -scan results (working with averages, etc.). So we will likely end up with a more robust detection by default AND adjustable sampling parameters.

[erwinkramer]

Some movies have just a couple of scenes with brightness expansion. I think 1 sample for each minute of playtime might be better, so that would be like about 100-150 samples.

[zhdenny]

Some movies have just a couple of scenes with brightness expansion. I think 1 sample for each minute of playtime might be better, so that would be like about 100-150 samples.

Ummm, that's all well and good until you are waiting hours for analysis to complete.
This is why it's best to be up to the user with sensible default

[erwinkramer]

@zhdenny yeah I didn't try to suggest that it would need to be the default, although it might be a nice additional strategy.

[cryptochrome]

@erwinkramer that's what we have the -inspect feature for. It analyses the entire movie, frame by frame. -scan is designed to give you a quick result, an overview. Taking 150 or more samples in -scan is probably going to be a lot slower than -inspect.

That being said: The "Complex FEL" detection in -scan is more or less fool-proof already. Typically, FELs that elevate luminance do this consistently across the movie (we use a very low 50 nit threshold). So taking 10 samples is already more than enough in most cases. The author of dovi_scripts (absolute authority on the subject) is currently working on a similar feature, and he will only take two samples. See the discussion here (Posts 1624, 1630 and 1631).

The biggest "threat" in Complex FEL detections is false positives. e. g. we flag something as Complex FEL because of a mastering error (single rogue pixel shooting to 10k nits). But better a false positive than a false negative.

As for Simple FEL: This is currently the biggest risk factor if we only take 10 samples. We might miss something, and that's why I currently advise to cross-check Simple FEL with -inspect.

All this will be improved with the new feature I am working on (adjustable number of scan probes and working with averages, like the dovi_scripts author).