More > JRiver Media Center 31 for Windows

Black bar detection not updating tag correctly

<< < (7/12) > >>

mattkhan:
my current feeling is that skip is the primary source of errors, in the absence of an auto option would it make sense to do something like run the analysis with a few different skip values and check the output is the same (for the same frames)? I suppose it would make the analysis x times slower but I think accuracy is preferred to speed in this case.

later I'll run a script against a load of films and see what it produces with different skip values

Hendrik:
Its not sorted right now, but I am sorting and filtering it for the upcoming changes. I can also increase the skip even further, and maybe investigate if there is options i can set to reduce the output of unfinished frames. Detecting if a frame is unfinished might not be possible, but I'll also check on that.

In this result, even without changes, the heuristic should pick the 1031 hit one with the proper dimensions, at least how I imagine it working. 47 hits for the full frame is just such a low percentage that it would be ignored.

mattkhan:
I wrote a little script https://github.com/3ll3d00d/jrmc-utils/blob/master/cropdetect.py which ploughs through my library, runs ffmpeg with a few difference parameters and dumps the result to some CSVs

I also wrote another script to summarise all of those into one csv https://github.com/3ll3d00d/jrmc-utils/blob/master/analyse_crop.py which compares vs the current mc video crop

the output of this one is like, I'm just printing the most popular single crop from ffmpeg here


--- Code: ---name,limit,skip,ffmpeg_crop,mc_crop
The Last Picture Show D1,18,24,0x0x3840x2160,0x0x3840x2160
The Last Picture Show D1,25,24,0x0x3840x2160,0x0x3840x2160
The Last Picture Show D1,18,8,0x0x3840x2160,0x0x3840x2160
The Last Picture Show D1,25,8,0x0x3840x2160,0x0x3840x2160
The Last Picture Show Original,18,24,0x0x3840x2160,0x42x3840x2118
The Last Picture Show Original,25,24,0x0x3840x2160,0x42x3840x2118
The Last Picture Show Original,18,8,0x0x3840x2160,0x42x3840x2118
The Last Picture Show Original,25,8,0x0x3840x2160,0x42x3840x2118
Your Name.,18,24,0x4x1920x1076,0x0x1920x1080
Your Name.,25,24,0x4x1920x1076,0x0x1920x1080
Your Name.,18,8,0x4x1920x1076,0x0x1920x1080
Your Name.,25,8,0x4x1920x1076,0x0x1920x1080
Malcolm X,18,24,0x4x1920x1076,0x0x1920x1080
Malcolm X,25,24,0x4x1920x1076,0x0x1920x1080
Malcolm X,18,8,0x4x1920x1076,0x0x1920x1080
Malcolm X,25,8,0x4x1920x1076,0x0x1920x1080
A Star Is Born,18,24,0x164x1920x916,0x0x1920x1080
A Star Is Born,25,24,0x164x1920x916,0x0x1920x1080
A Star Is Born,18,8,0x4x1920x1076,0x0x1920x1080
A Star Is Born,25,8,0x4x1920x1076,0x0x1920x1080

--- End code ---

I was thinking of then doing some sort of analysis on this like

* count how many agree then ignore those
* calculate the delta between the two crops, count how many have a "trivial" pixel difference then ignore those
* count how many are insensitive to limit/skip values
* try to scrape AR from something like blu-ray.com (though I don't think they have an API so that could get annoying), assess MC and ffmpeg vs that
* count how many, if any, have obviously wrong values
* if MC behaviour changes, assess how many videos were impacted by the change
* some other stuff I haven't thought of :)

I thought that this might provide a way to systematically compare MC's behaviour as it evolves across a large no of videos which might help arrive at an optimal solution more quickly

I can post the results once it completes (and if you want the raw data then I can share that also)

mattkhan:
ran that for ~600 films (ignored DVD), lots of differences but it's clear there are many trivial differences so I need to filter those out first

mattkhan:
hard to say if this is useful but I ran the analysis anyway on ~600 films :)

if I allow a tolerance of +/- 16 pixels on either dimension then I'm left with 74 BD or UHD with a significantly different crop to that produced by MC atm
if I reduce this tolerance to +/- 8 pixels then this number goes up dramatically (to ~250)
if I increase it (even to 32), the number of differences barely shifts so I think those 74 films are the material differences

of those 74, the ffmpeg crop is sensitive to

limit: 29
skip: 33
both: 13
neither: 25

i.e.

25 are insensitive to the values I chose for limit/skip
13 are sensitive to both (i.e. I get 4 different crops)
33 are sensitive to skip only (so I guess these are the ones with decoding errors)
29 are sensitive to the black level limit

reminds me that I need to further filter the differences within ffmpeg to check whether the different crops produces are trivial differences, probably this reduces the sensitivity somewhat (though not sure to what extent)

NB: there is at least 1 item (2001 4k) in which MC's current analysis is currently right and ffmpeg is not, not obvious why that is so probably need to investigate that one further to look at what ffmpeg was doing

Navigation

[0] Message Index

[#] Next page

[*] Previous page