@tritical
I wait for your next tivtc version, because I want to test you new three way mode.
So far so good, for today I stop with my test. I have learned a lot from testing the different sources. First of all my idea that after a clear field without a match can follows a second clear field without any possible match seems to be wrong. BUt my special mode for tfm I thought of wouldn´t be useless.
I made tests on ten different sources and more then eighty different motion scenes. On four of ten sources, there are sometimes (really seldom and only light) two blended fields, that stands together, so bad are the conversations.
It is possible, that if we have an u-match after an p-match, we lost a clear field.
This simply means, that we only should have a look to the next two fields after the last possible match. If we have a p-match and no possible c-match (there are maybe both matches possible, but p is better) we just should try to match the next two fields (second field from the previous frame and first field frame the frame we try to restore). That means only c- or p-match would be recommed. With a possible c-match of the previous frame also an u-match would be useful. The other thing is of course, that only this two fields should be used for the postprocessing with blend detection.
If we use tdeint as postprocessor after tfm we won´t solve this problem. Maybe you can think about the implemetation of such a mode in the feature.
But now to the more important thing, the blend detection tests.
Sorry tritical, but your idea also wouldn´t work that good. In nearly all scenes I looked at, I found some blends with blend factors between 10 and 15%, so the difference between MIN(0,4) and MIN(1,2,3) wouldn´t be that high or clear and sometimes we would detect the wrong field as blended.
A better method seems to be, that we compare the MIN(1,2,3) from testing the first field with the MIN(1,2,3) from testing the second field. In some cases this seems to work better, but also not for all absolut right.
I´m happy that my idea seems to work pretty good in all cases, I tested till now. To find the right blend factor by comparing the matches 0 and 4 seems to work really fine (till now not tested on really noisy and blocky source) also for the different sources. I still would prefer to compare the combing levels from the matched blend creations with the test of the other field. In some cases this seems to work better than compare the differences from MIN(0,4) with this combing level.
I made a list of the blending factors depends on the combing level ratio (0,4). I made it that accurate, clear because this should work better. Depening on the source the variance to the real blend factor isn´t higher than 3% (for my tests). I think this sounds not bad. I´ll make some more tests in the feature and maybe will modify the list a little bit, but till now this seems to work good. I hope these tests would help you a little bit.
So, here is the list
Code:
A = the ratio between the both matches with the blended field
(used the match values displayed from telecide)
B = the percentage of the blend factor from the field with the lower match values
A B A B
(value:1)
1 50% 2.5 29%
1.02 49% 2.6 28%
1.05 48% 2.7 27%
1.08 47% 2.9 26%
1.12 46% 3.0 25%
1.16 45% 3.2 24%
1.20 44% 3.4 23%
1.24 43% 3.6 22%
1.28 42% 3.8 21%
1.33 41% 4.1 20%
1.38 40% 4.4 19%
1.44 39% 4.7 18%
1.5 38% 5.0 17%
1.6 37% 5.4 16%
1.7 36% 5.8 15%
1.8 35% 6.3 14%
1.9 34% 6.7 13%
2.0 33% 7.2 12%
2.2 32% 7.7 11%
2.3 31% 8.2 10%
2.4 30%