Improving Madvr HDR to SDR mapping for projector - Page 200 - AVS Forum | Home Theater Discussions And Reviews
Forum Jump: 
 4753Likes
Reply
 
Thread Tools
post #5971 of 6029 Old 05-25-2019, 01:58 AM
AVS Forum Special Member
 
madshi's Avatar
 
Join Date: May 2005
Posts: 7,571
Mentioned: 456 Post(s)
Tagged: 0 Thread(s)
Quoted: 2328 Post(s)
Liked: 2987
Here's a modified Excel sheet to let you try your line1+2+3 suggestion:

http://madshi.net/NeoV19b.xlsx
Manni01, Fer15 and Neo-XP like this.
madshi is offline  
Sponsored Links
Advertisement
 
post #5972 of 6029 Old 05-25-2019, 04:12 AM
Advanced Member
 
Neo-XP's Avatar
 
Join Date: Jun 2018
Location: Switzerland
Posts: 835
Mentioned: 134 Post(s)
Tagged: 0 Thread(s)
Quoted: 572 Post(s)
Liked: 684
@madshi Thanks!

I combined both methods to the same sheet so it's easier to compare when adding new scenes:



Doc: https://www.mediafire.com/file/w41jk...oV20.xlsx/file

Line1 seems indeed not useful with my suggestion. I added more false positives, but it doesn't change the result.

Edit: You can get 1 less miss with build 85 if you switch Line1 with Line2 and then set Metric1 on Line1 to 65 or more.
Fer15 likes this.

Last edited by Neo-XP; 05-25-2019 at 05:39 AM.
Neo-XP is offline  
post #5973 of 6029 Old 05-25-2019, 04:59 AM
Senior Member
 
Join Date: Jan 2018
Posts: 350
Mentioned: 67 Post(s)
Tagged: 0 Thread(s)
Quoted: 188 Post(s)
Liked: 219
I decided to have a lie-in this morning. Looks like I missed all the fun!
Good work guys!

Last edited by Fer15; 05-25-2019 at 05:03 AM.
Fer15 is offline  
Sponsored Links
Advertisement
 
post #5974 of 6029 Old 05-25-2019, 06:07 AM
AVS Forum Special Member
 
madshi's Avatar
 
Join Date: May 2005
Posts: 7,571
Mentioned: 456 Post(s)
Tagged: 0 Thread(s)
Quoted: 2328 Post(s)
Liked: 2987
Quote:
Originally Posted by Neo-XP View Post
I combined both methods to the same sheet so it's easier to compare when adding new scenes:

Doc: https://www.mediafire.com/file/w41jk...oV20.xlsx/file
With bars this time, the sheet keeps getting prettier...

Quote:
Originally Posted by Neo-XP View Post
Line1 seems indeed not useful with my suggestion. I added more false positives, but it doesn't change the result.

Edit: You can get 1 less miss with build 85 if you switch Line1 with Line2 and then set Metric1 on Line1 to 65 or more.
You mean you can get to 20 misses that way? I've just tried and can't seem to get an improvement that way. 21 misses is still my best result.

Here's a new build:

http://madshi.net/madVRhdrMeasure86.zip

1) You need to enter all M1 and M2 thresholds multiplied by 10 now. So if you had entered 10 in the last build, now you have to enter 100. This allows more precision. E.g. entering "32" is now practically 3.2. Yeah, I know, I could have allowed true floating point edit controls instead, but that would have cost more development time, so I cheapened out by simply doing a 10x factor on the thresholds.

2) Line 1 is now "AND" operator instead of "OR".

3) Line 2 is gone. So the previous line 3 is now line 2.

4) Your M1/M2 settings are reset to the new defaults now, which are 65|15 for line 1, and 10|3.2 for line 2 (previously line 3).

Now the big question is if these default settings also work well in real life? Or are they only good for the Excel sheet, but make problems during real world playback?
Manni01 and NoTechi like this.
madshi is offline  
post #5975 of 6029 Old 05-25-2019, 06:30 AM
Advanced Member
 
Neo-XP's Avatar
 
Join Date: Jun 2018
Location: Switzerland
Posts: 835
Mentioned: 134 Post(s)
Tagged: 0 Thread(s)
Quoted: 572 Post(s)
Liked: 684
Quote:
Originally Posted by madshi View Post
You mean you can get to 20 misses that way? I've just tried and can't seem to get an improvement that way. 21 misses is still my best result.
No, sry, it was to avoid 1 miss with build 84, but in the end it was exactly like it is now with build 85

I can't get less than 21 either, even playing with M1 weight (it stays around 50% by trial and error).

Time to find new false positives, it won't be easy

Last edited by Neo-XP; 05-25-2019 at 06:57 AM.
Neo-XP is offline  
post #5976 of 6029 Old 05-25-2019, 07:03 AM
AVS Forum Special Member
 
madshi's Avatar
 
Join Date: May 2005
Posts: 7,571
Mentioned: 456 Post(s)
Tagged: 0 Thread(s)
Quoted: 2328 Post(s)
Liked: 2987
Quote:
Originally Posted by Neo-XP View Post
Time to find new false positives, it won't be easy
Hehe... How much time do you think we need for this stage of finding new problem scenes and fine tuning the M1/M2 thresholds etc?
madshi is offline  
post #5977 of 6029 Old 05-25-2019, 09:21 AM
Senior Member
 
Join Date: Jan 2015
Posts: 383
Mentioned: 8 Post(s)
Tagged: 0 Thread(s)
Quoted: 292 Post(s)
Liked: 160
Hi, if outputting as HDR, are these test builds able to mimic dynamic tonemapping as found in some TVs? It all looks quite complicated, how would i set this up? I have an LG C8, and while the dynamic tonemapping feature works ok, it does tend to overly brighten things.
iSeries is online now  
post #5978 of 6029 Old 05-25-2019, 12:40 PM
Advanced Member
 
ddgdl's Avatar
 
Join Date: Mar 2007
Posts: 606
Mentioned: 6 Post(s)
Tagged: 0 Thread(s)
Quoted: 328 Post(s)
Liked: 298
Quote:
Originally Posted by iSeries View Post
Hi, if outputting as HDR, are these test builds able to mimic dynamic tonemapping as found in some TVs? It all looks quite complicated, how would i set this up? I have an LG C8, and while the dynamic tonemapping feature works ok, it does tend to overly brighten things.
Wrong thread for that- go to the madvr player help thread
ddgdl is offline  
post #5979 of 6029 Old 05-25-2019, 12:40 PM
Advanced Member
 
Neo-XP's Avatar
 
Join Date: Jun 2018
Location: Switzerland
Posts: 835
Mentioned: 134 Post(s)
Tagged: 0 Thread(s)
Quoted: 572 Post(s)
Liked: 684
Quote:
Originally Posted by madshi View Post
Hehe... How much time do you think we need for this stage of finding new problem scenes and fine tuning the M1/M2 thresholds etc?
Not much, but I'm wondering if it wouldn't be better to remove all false positives from flashes from the sheet.

Actually, unlike false positives from movements, they do not produce visible artifacts at all, and better than this, letting them detected as new scenes improves the image quality most of the time (getting to the ideal target transparently).

For instance:

Before flash / Flash ignored / Flash detected as new scene



Thoughts?
Neo-XP is offline  
post #5980 of 6029 Old 05-25-2019, 12:46 PM
Senior Member
 
Join Date: Jan 2018
Posts: 350
Mentioned: 67 Post(s)
Tagged: 0 Thread(s)
Quoted: 188 Post(s)
Liked: 219
Quote:
Originally Posted by Neo-XP View Post
Not much, but I'm wondering if it wouldn't be better to remove all false positives from flashes from the sheet.

Actually, unlike false positives from movements, they do not produce visible artifacts at all, and better than this, letting them detected as new scenes improves the image quality most of the time (getting to the ideal target transparently).

For instance:

Before flash / Flash ignored / Flash detected as new scene



Thoughts?

Agreed that scene looks better, but how would you distinguish that flash scene from the gunfire muzzle flash in BvS (which looks terrible)?
Fer15 is offline  
post #5981 of 6029 Old 05-25-2019, 12:50 PM
Advanced Member
 
Neo-XP's Avatar
 
Join Date: Jun 2018
Location: Switzerland
Posts: 835
Mentioned: 134 Post(s)
Tagged: 0 Thread(s)
Quoted: 572 Post(s)
Liked: 684
Quote:
Originally Posted by Fer15 View Post
Agreed that scene looks better, but how would you distinguish that flash scene from the gunfire muzzle flash in BvS (which looks terrible)?
This one last for a few seconds, the ones from BvS only for 1 or 2 frames (they can be removed easily when looking ahead).

When the flash goes ON and OFF quickly (few frames), I think it's better not to adapt the target at all, but not there.
Fer15 likes this.

Last edited by Neo-XP; 05-25-2019 at 12:56 PM.
Neo-XP is offline  
post #5982 of 6029 Old 05-25-2019, 12:51 PM
AVS Forum Special Member
 
madshi's Avatar
 
Join Date: May 2005
Posts: 7,571
Mentioned: 456 Post(s)
Tagged: 0 Thread(s)
Quoted: 2328 Post(s)
Liked: 2987
In any case, would it change anything if you were to remove those flashes from the Excel sheet? Would the optimal thresholds and weights change? Much?
madshi is offline  
post #5983 of 6029 Old 05-25-2019, 01:35 PM
Senior Member
 
Join Date: Jan 2018
Posts: 350
Mentioned: 67 Post(s)
Tagged: 0 Thread(s)
Quoted: 188 Post(s)
Liked: 219
Quote:
Originally Posted by madshi View Post
In any case, would it change anything if you were to remove those flashes from the Excel sheet? Would the optimal thresholds and weights change? Much?

Just by quickly sorting the results, it looks like flashes in the false positive section generally have a larger metric 2 result than movements.


Average metric 2 false positive for flash = 4.98947368


Average metric 2 false positive for movement =1.09611111


Average metric 2 false positive for flash and movement = 1.121


Haven't altered the thresholds as Neo is working on some new false positives atm I believe.
Fer15 is offline  
post #5984 of 6029 Old 05-25-2019, 02:12 PM
Advanced Member
 
Neo-XP's Avatar
 
Join Date: Jun 2018
Location: Switzerland
Posts: 835
Mentioned: 134 Post(s)
Tagged: 0 Thread(s)
Quoted: 572 Post(s)
Liked: 684
Quote:
Originally Posted by madshi View Post
In any case, would it change anything if you were to remove those flashes from the Excel sheet? Would the optimal thresholds and weights change? Much?
Yes, Fer15 sent me a lot of false positives from flashes only, and as adding more of them to the sheet, the thresholds need to be increased to keep the lowest sum of misses.

Of course, some real scenes will be missed in favor of false positives from flashes (that are not producing any artifacts) if I continue to add scenes like this. So do I continue this way?

Doc: https://www.mediafire.com/file/2hl7l...oV21.xlsx/file
Manni01 and Fer15 like this.
Neo-XP is offline  
post #5985 of 6029 Old 05-25-2019, 02:17 PM
AVS Forum Special Member
 
madshi's Avatar
 
Join Date: May 2005
Posts: 7,571
Mentioned: 456 Post(s)
Tagged: 0 Thread(s)
Quoted: 2328 Post(s)
Liked: 2987
I don't know. Do those real scenes which are then missed produce visible artifacts?
madshi is offline  
post #5986 of 6029 Old 05-25-2019, 02:34 PM
Advanced Member
 
Neo-XP's Avatar
 
Join Date: Jun 2018
Location: Switzerland
Posts: 835
Mentioned: 134 Post(s)
Tagged: 0 Thread(s)
Quoted: 572 Post(s)
Liked: 684
@madshi No, not at all for now.
Neo-XP is offline  
post #5987 of 6029 Old 05-25-2019, 11:56 PM
AVS Forum Special Member
 
madshi's Avatar
 
Join Date: May 2005
Posts: 7,571
Mentioned: 456 Post(s)
Tagged: 0 Thread(s)
Quoted: 2328 Post(s)
Liked: 2987
Ok, so we have the choice to optimize the thresholds for either avoiding flash false positives better, or for detecting real scenes better, but neither misses seem to produce visible artifacts atm. I wonder how often do those flash situations appear in real movies? And how often do real scenes appear which are hard to detect for madVR? I don't know for sure, but I'd tend to say that those difficult to detect real scenes are more common? In which case we probably shouldn't add *too* many flash scenes to the Excel sheet.

Also, there's an argument to be made that if a flash scene stays for more than 2-3 frames, then although it's strictly speaking not a different scene, from the lighting (not lightning) point of view, it does sort of classify as a completely different scene, so adjusting the tone mapping immediately might actually make sense. So I tend to prefer not putting too much weight on such flash scenes right now.

One thing that's confusing me a bit, though: Isn't a flash supposed to be very short? How come it's longer than 2-3 frames? FWIW, I was planning to look ahead maybe 3 frames, but I'm not sure yet. If such flash scenes were shorter than that, the looking ahead should already make sure they're not detected as a scene change.
Neo-XP likes this.
madshi is offline  
post #5988 of 6029 Old 05-26-2019, 04:35 AM
Senior Member
 
Join Date: Jan 2018
Posts: 350
Mentioned: 67 Post(s)
Tagged: 0 Thread(s)
Quoted: 188 Post(s)
Liked: 219
I'm open-minded about whether to register a flash that lasts longer than a few frames as a new scene. The Atomic Blonde example looks better imo. And looking ahead to discard intermittent flashing for a frame or two would avoid the BvS gunfire muzzle flash....but what if that guy held the trigger down for a few seconds? Then that would be similar to Atomic Blonde and you'd get that noticeable dimming before gunfire/after gunfire, which would be noticeable if it were to be registered as a new scene imo,.

So I'm wondering if there were another way to distinguish a flash in addition to its duration i.e. to impose a limiter of some sort that wouldn't activate if the flash was longer in BvS?

BvS:

goes from measured frame peak 400 nits frame FALL around 50 to 5000 nits at 800 frame FALL

and the second instance of muzzle fire goes from measured frame peak 100 nits frame FALL around 4 to 5400 nits at 400 frame FALL (although I don't think this is as noticeable as the previous scene above)


For Atomic Blonde:

goes from measured frame peak 175 nits frame FALL around 40 to 300 nits at 110 frame FALL


However, I've just looked at the metric stats

BvS = 29.19 / 4.14 / 2.11

Atomic Blonde = 75.66 / 16.02 / 6.29

That's quite a difference, and Atomic Blonde might have high enough stats to be registered as a new scene in most circumstances under the proposed thresholds. And BvS might fall under the threshold.
Fer15 is offline  
post #5989 of 6029 Old 05-26-2019, 04:36 AM
Advanced Member
 
Neo-XP's Avatar
 
Join Date: Jun 2018
Location: Switzerland
Posts: 835
Mentioned: 134 Post(s)
Tagged: 0 Thread(s)
Quoted: 572 Post(s)
Liked: 684
Quote:
Originally Posted by madshi View Post
One thing that's confusing me a bit, though: Isn't a flash supposed to be very short? How come it's longer than 2-3 frames? FWIW, I was planning to look ahead maybe 3 frames, but I'm not sure yet. If such flash scenes were shorter than that, the looking ahead should already make sure they're not detected as a scene change.
Maybe the word "flash" is not right, it's more sudden lightning changes from one frame to the next. When someone turns on the lights, it's not really a flash, but there is no difference for the current detection.

It can happen for a few frames (1-3 max) with strobe lights (nightclubs, police lights), flashlight (pointing to camera), fires, sun (e.g. through the windows of a driving car), muzzle flashes, etc.

After removing those cases from the sheet, I get coherent results again:



Metric1 on Line1 needs to be set to 100, because when Metric1 fails (bar on the histogram moving left/right), it can be as high as 91.72, while Metric2 is near 0.

Metric2 at 1.6 on Line1 produces the least number of overall misses.

Metric1 at 13.4 and Metric2 at 3.5 give me the best detection ratio and excellent real scenes detection too.

So this is it (for me):


Edit: I forgot the doc: https://www.mediafire.com/file/qpzdx...oV22.xlsx/file

Edit2:

Quote:
Originally Posted by Fer15 View Post
but what if that guy held the trigger down for a few seconds?
If the FALL doesn't go up and then down after a few frame, it's probably better just to detect it as a new scene, but I've never seen a case like this (it always goes up and down after 1 or max 2 frames on muzzle flashes).

Quote:
Originally Posted by Fer15 View Post
BvS = 29.19 / 4.14 / 2.11

Atomic Blonde = 75.66 / 16.02 / 6.29

That's quite a difference, and Atomic Blonde might have high enough stats to be registered as a new scene in most circumstances under the proposed thresholds. And BvS might fall under the threshold.
Other muzzle flashes goes higher than this in BvS.

For instance at frame 196772:

43.69 / 5.79 / 2.46

How many real scenes will we miss just to ignore this false positive? a lot.

However, if you can just look ahead for 1-2 frame(s), you can discard 99.99% of those without hurting the real scenes detection.
Manni01 and Fer15 like this.

Last edited by Neo-XP; 05-26-2019 at 05:16 AM.
Neo-XP is offline  
post #5990 of 6029 Old 05-26-2019, 05:54 AM
Senior Member
 
Join Date: Jan 2018
Posts: 350
Mentioned: 67 Post(s)
Tagged: 0 Thread(s)
Quoted: 188 Post(s)
Liked: 219
Quote:
Originally Posted by Neo-XP View Post

Edit2:



If the FALL doesn't go up and then down after a few frame, it's probably better just to detect it as a new scene, but I've never seen a case like this (it always goes up and down after 1 or max 2 frames on muzzle flashes).


Other muzzle flashes goes higher than this in BvS.

For instance at frame 196772:

43.69 / 5.79 / 2.46

How many real scenes will we miss just to ignore this false positive? a lot.

However, if you can just look ahead for 1-2 frame(s), you can discard 99.99% of those without hurting the real scenes detection.


Yeah, I guess that's true. I've just looked through some firing scenes i.e. Rambo, MI: Fallout; Saving Private Ryan; Die Hard and those are only intermittent frames for muzzle fire.


The mini-gun scene from T2 can be longer in duration (4 consecutive frames in one sequence circa 140350 and five frames from around 140795) (although it's a dim film, some of the muzzle fire is slight and the dynamic tone mapping doesn't fluctuate much)


Of course, this is just for one gun. I guess the same effect could happen with a scene of multiple instances of gunfire going of simultaneously.


edit: also the mini gun sequence from Predator from around frame 69143 has a consecutive muzzle fire flash of 30-40 frames (but these aren't creating a false positive)


I don't believe it's beyond the realm of possibility that some might be as long (longer than 4/5 fames) with big changes in fameFALL similar to BvS (although BvS is probably on the more extreme end with very high nits for muzzle fire) but until we can find any specific examples, I guess it's a moot point..
Neo-XP likes this.

Last edited by Fer15; 05-26-2019 at 09:17 AM.
Fer15 is offline  
post #5991 of 6029 Old 05-26-2019, 06:10 AM
Advanced Member
 
Neo-XP's Avatar
 
Join Date: Jun 2018
Location: Switzerland
Posts: 835
Mentioned: 134 Post(s)
Tagged: 0 Thread(s)
Quoted: 572 Post(s)
Liked: 684
Another sequence for which is better not to ignore the false new scene:

New scene ignored




New scene detected




Even for just 2 frames:

New scene ignored




New scene detected



I don't know if it will be possible to distinguish this from muzzle flashes when looking ahead, but that is for another time (when madshi takes the DeLorean ).

PS: I'm done with build 86.
Neo-XP is offline  
post #5992 of 6029 Old 05-26-2019, 06:38 AM
Senior Member
 
Join Date: Jan 2018
Posts: 350
Mentioned: 67 Post(s)
Tagged: 0 Thread(s)
Quoted: 188 Post(s)
Liked: 219
Quote:
Originally Posted by Neo-XP View Post
Another sequence for which is better not to ignore the false new scene:

New scene ignored




New scene detected




Even for just 2 frames:

New scene ignored




New scene detected



I don't know if it will be possible to distinguish this from muzzle flashes when looking ahead, but that is for another time (when madshi takes the DeLorean ).

.









Found another scene that would benefit from looking ahead for intermittent muzzle flashes



The Matrix


Manni01, omarank and Neo-XP like this.
Fer15 is offline  
post #5993 of 6029 Old 05-26-2019, 10:36 AM
AVS Forum Special Member
 
madshi's Avatar
 
Join Date: May 2005
Posts: 7,571
Mentioned: 456 Post(s)
Tagged: 0 Thread(s)
Quoted: 2328 Post(s)
Liked: 2987
So what happens if two frames are ultra bright and the frames around them are dark? If I look ahead, should I generally treat those 2 bright scenes as a different frame? Or should I not? If it's only one frame I should most probably not detect a scene change, right? But how many frames does it take to justify treating them as a scene change?
madshi is offline  
post #5994 of 6029 Old 05-26-2019, 11:19 AM
Advanced Member
 
Neo-XP's Avatar
 
Join Date: Jun 2018
Location: Switzerland
Posts: 835
Mentioned: 134 Post(s)
Tagged: 0 Thread(s)
Quoted: 572 Post(s)
Liked: 684
Quote:
Originally Posted by madshi View Post
So what happens if two frames are ultra bright and the frames around them are dark? If I look ahead, should I generally treat those 2 bright scenes as a different frame? Or should I not? If it's only one frame I should most probably not detect a scene change, right? But how many frames does it take to justify treating them as a scene change?
I'd say it doesn't depend only on the number of frames, but also on the context.

When the FALL goes way higher for 1 or a few frames between two dark scenes:

- If only a (small) part of the image gets way brighter, it seems better not to adapt the target.

- If the entire image gets way brighter, it seems better to adapt to "ideal" target every time, even just for 1 frame.

Maybe there is a way to use the lum histogram for this...

Last edited by Neo-XP; 05-26-2019 at 11:36 AM.
Neo-XP is offline  
post #5995 of 6029 Old 05-26-2019, 11:45 AM
AVS Forum Special Member
 
madshi's Avatar
 
Join Date: May 2005
Posts: 7,571
Mentioned: 456 Post(s)
Tagged: 0 Thread(s)
Quoted: 2328 Post(s)
Liked: 2987
Hmmmm... Yes, I suppose that makes sense. We'll have to test that in more detail when I have looking ahead implemented.

So for now I'll use 1000, 16, 134, 35, 50 as new defaults, right?
madshi is offline  
post #5996 of 6029 Old 05-26-2019, 11:48 AM
Senior Member
 
Join Date: Jan 2018
Posts: 350
Mentioned: 67 Post(s)
Tagged: 0 Thread(s)
Quoted: 188 Post(s)
Liked: 219
Quote:
Originally Posted by madshi View Post
So what happens if two frames are ultra bright and the frames around them are dark? If I look ahead, should I generally treat those 2 bright scenes as a different frame? Or should I not? If it's only one frame I should most probably not detect a scene change, right? But how many frames does it take to justify treating them as a scene change?



If it's one frame only, I'd ignore it and not regard it as a scene change. One frame would not be perceptible to the viewer in real time. I had a look at that BvS scene (the second image which Neo posted above that was only 2 frames), but in real time, that wasn't perceptible imo.

My initial impression is that anything that is four/ five frames or less, should not be treated as a scene change.


I'll need to look at other scenes before being definitive about it.



edit: maybe science about human visual perception could guide us?


quick google search:



Bloch's law. “Basically, it’s one of the few laws in perception,” Professor Thomas Busey, associate department chair at Indiana University’s Department of Psychological and Brain Sciences, tells me. It says that there’s a trade-off between intensity and duration in a flash of light lasting less than 100ms. You can have a nanosecond of incredibly bright light and it will appear the same as a tenth of a second of dim light. “In general, people can’t distinguish between short, bright and long, dim stimuli within a tenth of a second duration,”



and 13ms for image recognition according to this link https://link.springer.com/article/10...414-013-0605-z


Quote:
Originally Posted by Neo-XP View Post
I'd say it doesn't depend only on the number of frames, but also on the context.

When the FALL goes way higher for 1 or a few frames between two dark scenes:

- If only a small part of the image gets way brighter, it seems better not to adapt the target.

- If the entire image gets way brighter, it seems better to adapt to "ideal" target every time, even just for 1 frame.

Maybe there is a way to use the lum histogram for this...

My concern about the latter ("If the entire image gets way brighter, it seems better to adapt to "ideal" target every time, even just for 1 frame") is that the likes of a close-up muzzle flash i.e. BvS, The Matrix might constitute the "entire image"and even resetting it for one frame would cause a noticeable dimming i.e. before gunfire/ after gunfire.

Last edited by Fer15; 05-26-2019 at 12:25 PM.
Fer15 is offline  
post #5997 of 6029 Old 05-26-2019, 12:44 PM
Advanced Member
 
Neo-XP's Avatar
 
Join Date: Jun 2018
Location: Switzerland
Posts: 835
Mentioned: 134 Post(s)
Tagged: 0 Thread(s)
Quoted: 572 Post(s)
Liked: 684
Quote:
Originally Posted by madshi View Post
So for now I'll use 1000, 16, 134, 35, 50 as new defaults, right?
Yes, that's pretty good. Only 18 misses on 290 scenes

And it works great on real conditions too.

Quote:
Originally Posted by Fer15 View Post
My concern about the latter ("If the entire image gets way brighter, it seems better to adapt to "ideal" target every time, even just for 1 frame") is that the likes of a close-up muzzle flash i.e. BvS, The Matrix might constitute the "entire image"and even resetting it for one frame would cause a noticeable dimming i.e. before gunfire/ after gunfire.
BvS and The Matrix scenes are good examples of only a part of the image (the muzzle flashes) getting way brighter:



Even in close-up:





The histogram changes seem quite different when the entire image is affected:





But I don't know, maybe there is another way to identify them...
Manni01 and Fer15 like this.

Last edited by Neo-XP; 05-26-2019 at 12:53 PM.
Neo-XP is offline  
post #5998 of 6029 Old 05-26-2019, 01:18 PM
Senior Member
 
Join Date: Jan 2018
Posts: 350
Mentioned: 67 Post(s)
Tagged: 0 Thread(s)
Quoted: 188 Post(s)
Liked: 219
Quote:
Originally Posted by Neo-XP View Post
Yes, that's pretty good. Only 18 misses on 290 scenes

And it works great on real conditions too.



BvS and The Matrix scenes are good examples of only a part of the image (the muzzle flashes) getting way brighter:


Even in close-up:


The histogram changes seem quite different when the entire image is affected:



But I don't know, maybe there is another way to identify them...

If that's enough for Madshi to work with to distinguish those types of images then great!
Fer15 is offline  
post #5999 of 6029 Old 05-26-2019, 02:03 PM
Senior Member
 
Join Date: Jan 2018
Posts: 350
Mentioned: 67 Post(s)
Tagged: 0 Thread(s)
Quoted: 188 Post(s)
Liked: 219
There's also this scene from Mad Max, which is around 4 frames long (frame 40712 or there about)

(big spike on the lum chart)


before/ false positive (incorrect scene change) / no false positive








difference in dynamic tone mapping. Around 250 vs 1500 + vs 250


oh and btw, there are a few more false positives for muzzle fire in Predator at frame 36412 and 109331 (single frame instance like BvS)
Manni01 likes this.

Last edited by Fer15; 05-26-2019 at 02:47 PM.
Fer15 is offline  
post #6000 of 6029 Old 05-27-2019, 04:33 AM
Advanced Member
 
Neo-XP's Avatar
 
Join Date: Jun 2018
Location: Switzerland
Posts: 835
Mentioned: 134 Post(s)
Tagged: 0 Thread(s)
Quoted: 572 Post(s)
Liked: 684
Quote:
Originally Posted by Fer15 View Post
oh and btw, there are a few more false positives for muzzle fire in Predator at frame 36412 and 109331 (single frame instance like BvS)
They seem "easy" to detect, based on the histogram changes:



Neo-XP is offline  
Sponsored Links
Advertisement
 
Reply Digital Hi-End Projectors - $3,000+ USD MSRP

Tags
dynamic tone mapping , hdr , madvr , sdr , ton mapping

Thread Tools
Show Printable Version Show Printable Version
Email this Page Email this Page


Forum Jump: 

Posting Rules  
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

BB code is On
Smilies are On
[IMG] code is On
HTML code is Off
Trackbacks are Off
Pingbacks are Off
Refbacks are Off