Use post=4 (and post=5) rather than post=2 (and post=3). Also, do not set an unnecessarily low vthresh parameter.
The dthreshold parameter was renamed to dthresh in Telecide for Decomb 5.0.0. It is still called dthreshold in FieldDeinterlace.
It's faster and less error-prone to do 2 comparisons rather than 3. But Decomb 5.0.0 will still try the third match if a frame is still combed after the 2-way match. So no extra combed frames are emitted compared to Decomb 4. The tutorial provides a really easy way to determine a clip's field order.
First, verify that your field order is set correctly using the procedure given in the tutorial!
The field matching process is not perfect. However, it can be improved in some situations. For example, if you have a noisy clip (as you might get from an analog capture), you can often improve the matching by using the nt parameter to add more noise tolerance than the default value of 10. The best value will depend on how noisy your clip is. But be cautious when raising the noise tolerance, because it can degrade matching for clean clips.
You can also improve matching in some cases by enabling pattern guidance. In some uncommon cases you can improve matching by disabling pattern guidance! Try it both ways and see which gives you better results.
These scene changes require a third (backward) match. Decomb will test the third match only if post is set to any value other than 0. Note that you can enable the third match testing without enabling postprocessing: you can set post=1. You can also enable three-way matching on all frames (not just ones that are combed after two-way matching) by using the triple parameter. Setting triple=true, however, can produce jerkiness.
First, use post=3 or post=5 to verify that the frame is actually being deinterlaced. If it is not, then you need to reduce your vthresh parameter until it is. If it is being deinterlaced, then you can decrease the dthresh parameter to attack more areas of the frame, but see the caveat below.
Make sure that what you think is combing is really combing. For example, sometimes the clip has content that is intentional but looks like combing. If you use SeparateFields() (by itself without Decomb) and still see "combing" on the fields, then it is not really combing but intentional content.
If you have a color conversion before Decomb, be sure to use ConvertToXXX(interlaced=true).
Do not resize vertically before Decomb.
Caveat About Residual Combing: People will often freeze on an frame and zoom in closely and see some faint residual combing and agonize over it. But that is a seriously misguided way to assess residual combing. Human perception cannot register contrast when the difference becomes less than a threshold. But the threshold gets lower as the feature size gets larger. That means that when you zoom in, you get an incorrect idea of what can actually be seen at normal size. Also, when things are in motion we see them with reduced resolution; therefore, freezing the frame gives an incorrect idea of what can be perceived. To assess residual combing, play the clip at normal speed and size. If it looks good, then be happy!
Finally, reducing dthresh to low values (such as 0-5) throws away most of the advantages of the area-based adaptive deinterlacing that Decomb performs. You will lose resolution on static areas of the frames, where you want it the most.
This artifact is a consequence of the field-differencing algorithm employed by FieldDeinterlace. All deinterlacers have characteristic faults like this. There is no perfect solution to deinterlacing. This is usually seen at the edges of letterboxing. You can crop off the letterboxes first and restore them afterwards (if you need them) as a workaround. Alternatively, you can use a frame-differencing deinterlacer, such as TomsMoComp, if you can accept the characteristic artifacts of frame differencing.
I am exploring ways to reduce this artifact and hope for better performance in the future.
The process goes like this (if full = false): First the frame is tested to see if it is combed. This is done by detecting combed areas and seeing if there is enough combing to reasonably declare the frame as combed. The threshold parameter determines the sensitivity of the combing detection in this step. Second, if the frame is declared as combed, the combed areas of the frame are again detected but this time using the dthreshold parameter. The areas of the frame detected as combed using dthreshold are then deinterlaced.
If full=true, the first step is omitted and all frames are declared as combed.
FieldDeinterlace's map parameter allows you to directly visualize the effect of these parameters.
Refer to the question on blending versus interpolation. To get rid of ghosting, set blend=false.
Suppose you have 3 pixels on successive lines and the same x offset:
a b cIf you decide that pixel b is combed, and you replace it with (0.5*a+0.5*c) that is (linear) interpolation. If you replace it with (0.25*a+0.5*b+0.25*c), that is blending. Note that when interpolating, the even lines are passed through and the odd lines are calculated as described. For blending, every line is calculated. Blending thus produces a blend of the two fields while interpolation uses data from one field only. That is why people say that blending causes "ghosting". You can eliminate the ghosting by setting blend=false. Also note that while interpolation can be calculated in place, blending cannot. Interpolation is a little faster.
The combing detector used by Decomb is quite sensitive but that can sometimes cause you trouble when you have content that has sections that are not combed but have detail that looks like combing. If you set vthresh high to avoid deinterlacing these areas, you may let combs slip through elsewhere. While it's better to err on the side of caution and deinterlace some good frames, some extreme cases make you wish you could set different vthresh values for different parts of the clip. Fortunately, Decomb allows this through its manual override capability. Refer to the reference manual for further details.
If the pattern guidance thinks that the calculated (blind) field match is not what it predicts based on pattern analysis, it tries to override the blind field match. The override is allowed if the pattern mismatch is less than gthresh. Think of gthresh as a limiter: it limits how large the mismatch can be before a predicted match is overruled by the actual calculated match. If your gthresh is too high, it may cause a new pattern in the clip to be erroneously overridden by the previous pattern. There is a happy medium that allows new patterns to be locked onto quickly while still allowing the pattern guidance to correct mismatches due to noise, etc. The happy medium is usually in the range 5-10%, but it will be clip-specific.
Copyright © Donald A. Graft, All Rights Reserved.
For updates and other filters/tools, visit my web site:
http://shelob.mordor.net/dgraft/
$Date: 2005/10/14 13:31:04 $