How do I use DeBlock with or without McTemporalDenoise?

[VideoFanatic]

Jagabo do you want to test it yourself here to see if McTemporalDenoise is actually deblocking properly or not? This is my script and I think I've set deblocking to full strength. I encoded the video as h264.

Mpeg2Source("L:\Raw 2001 Dec 10 stutter & brightness changing.d2v")

RemoveSpots()
AddGrainC(var=1.0,uvar=1.0)

McTemporalDenoise(settings="medium", interlaced=true, deblock=true, useQED=false, quant1=40, quant2=80)

ColorYUV(autogain=true)

Tweak(Bright=0, Sat=1.0, Cont=1.0, Hue=0.0, Coring=False)

Crop(22,0,-6,-12)
AddBorders(14,6,14,6, $000000)

If there's a problem then is the only way to fix it by changing the crop? What values are allowed? I know you said that Deblocking works on 8x8 blocks of pixels but that doesn't really mean anything to me.

I added AddgrainC in the location I did because I found that if I didn't have AddGrainC then I could still see spots on the stills in the video below. I watched it on my TV to see the difference. Since I don't actually want grain I added the denoiser after AddGrainC. I can't actually see grain but I figure I don't want it hence the denoiser. After that, AddGrainC still covers up the spots in the stills. I don't understand this myself but that's what I saw on my TV.

I've got thousands of videos. I can't check them all to see if I need RemoveSpots or not and I figure it doesn't do any harm so I always use it now.

Remove Spots from my video thread

In the thread you're reading now, McTemporalDenoise IS doing something, it's denoising which is what I wanted. I know Tweak isn't doing anything as I found the brightness, contrast, etc to be fine, I always keep Tweak there if I need it or not. Of course Sanlyn's script in the fix changing brightness levels thread is a revelation, I was just looking for a way to fix the changing levels of brightness but what I got was a lot more and it improved the video massively. I didn't even realise there were problems with the video (apart from changing levels) until I saw the difference that Sanlyn's script made. But that's best discussed in that thread. This thread is only for discussing deblocking with or without McTemporalDenoise.

[themaster1]

the code i use to deblock interlaced sources (gently provided by didée)

par=getparity()
SeparateFields().PointResize(width,height)
Deblock_QED(24,28,uv=3)
AssumeFrameBased()
SeparateFields()
Merge(SelectEven(),SelectOdd())
par ? AssumeTFF() : AssumeBFF()
Weave()

[jagabo]

Tonight I learn that any crop value not divisible by 8 is a no-no.

Only if you are going to deblock after cropping.

[VideoFanatic]

But I deblocked before cropping and you said I couldn't do that either?

[VideoFanatic]

Thanks poisondeathray and themaster. I'll try what you suggested.

[jagabo]

The better the deinterlacing before deblocking, the better the results will be.

[jagabo]

But I deblocked before cropping and you said I couldn't do that either

No, I didn't say that.

[sanlyn]

Tonight I learn that any crop value not divisible by 8 is a no-no.

Only if you are going to deblock after cropping.

Sorry, my Crop remarks do look kinda facetious. I get punchy by 2-AM. But I know what you're saying.

F347-original-interlaced (below, unmodified capture from the source mpg):
There are obvious macroblocks on the big guy's bare leg, near the knee, and one at chest level. Some less obvious blocks along the left border and along the length of the diagonal red rope. Clumped-up grainy stuff with chroma noise in dark areas and around the white lights on the right. Then there's color shift and bleed, mostly oversaturated red. Also, note the white ghost trails of the whitish figure on the right. Strong interlace combing and poor motion control. Most of this junk isn't "noise", it's low-bitrate artifacts and the usual tape playback problems.

[Attachment 14244 - Click to enlarge]


F347-MCTD-interlaced
McTemporalDenoise(settings="medium", interlaced=true, deblock=true, useQED=false, quant1=40, quant2=80)
Macroblocks softened a bit but still there. Clumpy grain smoothed a little, fine detail in the audience is disappearing. Edges visibly cleaner. Basically, the image has been smoothed but most artifacts are still there.

[Attachment 14245 - Click to enlarge]


F347_MCTD_Bob
Same MCTD (interlaced=false) but Bob'd first, then reinterlaced. Results similar to above, but more fine detail is retained.

[Attachment 14246 - Click to enlarge]

More in next post. . . .

[sanlyn]

F347_QTGMC_MCTD
QTGMC instead of Bob, same MCTD as earlier. Better detail, otherwise the blocks show few differences.

[Attachment 14247 - Click to enlarge]

F347_Bob_Deblock (progressive)
Interleave(Deblock(60).Bob(), Bob.Deblock(60))
I took the suggested script literally. Some smoothing, but distortion and odd contouring. It would take a convoluted reinterlace statement to keep it from running at 120 fps. I didn't have time to concoct a reinterlace.

[Attachment 14248 - Click to enlarge]


F347_Resize_QED.psd
par=getparity()
SeparateFields().PointResize(width,height)
Deblock_QED()
AssumeFrameBased()
SeparateFields()
Merge(SelectEven(),SelectOdd())
par ? AssumeTFF() : AssumeBFF()
Weave()
Again, not a quantum leap. Kinda deja-vu. And I'm beginning to think those blocks won't go away, ever.

[Attachment 14249 - Click to enlarge]

Next post . . .

[jagabo]

F347-original-interlaced (below, unmodified capture from the source mpg):
There are obvious macroblocks on the big guy's bare leg, near the knee, and one at chest level.

If you Bob() you'll see that both fields are full of macroblocks. You'll also see that all the even fields have screwed up chroma (a mix of the field before and the field after).

[sanlyn]

In my own script in the other thread of this same video, I didn't address the blocks in these frames. For one thing, they don't occur often; the action is so fast, no one sees this stuff. I didn't see lots of blocks elsewhere, but I'm sure there are more. But there's a reason I gave up on this particular issue, as follows:

F347_QTGMC_DeBlock
QTGMC(preset="slow")
DeBlock(quant=60,aOffset=8,bOffset=8)
Most of the lesser blocking looks clean, but the big blocks are now big blurs and contours remain distorted. Images in which this distortion occurs illustrate the basic problem: the encoder didn't just create macroblocks, it literally deformed parts of the image.



F347_Bob_Even (top image) and F347_Bob_Odd (lower)
Looking at bob'd or separated fields, you can see the distorted contours the deblockers can't fix. You might be able to play with motion interpolation here, but it would involve many fast-action frames and would take forever.




F346_Bob_Odd
On top of that, there aren't that many "clean" fields to work with. Take a look at the preceding frame 346, where the distortion is just as bad if not worse (look at the knee):



You see this kind of garbage all the time with inappropriate bitrates, yet many viewers stubbornly over-cram lots of video into small spaces and insist that the results "look great". They don't. Not by a long shot.

[jagabo]

Interleave(Deblock(60).Bob(), Bob.Deblock(60))
I took the suggested script literally. Some smoothing, but distortion and odd contouring. It would take a convoluted reinterlace statement to keep it from running at 120 fps.

That wasn't suggested as a method of deblocking the video. It was so he could compare what happens when you try to deblock an interlaced video, vs deinterlacing before deblocking. You open the script in VirtualDub and use the arrow keys to flip back and forth between the two videos.

One field of the original video (simple Bob()):


Deblock(60).Bob():


Bob().Deblock(60):


If you zoom in on the original video you'll see that most of the frame has blocky artifacts. The frame that was deblocked while interlaced has some of the blocks smoothed out but has more jagged edges and further screwed up chroma channels. The frame that was bobbed before deblocking has most of the blocks smoothed away and the chroma isn't contaminated with chroma from the other field.

Again, this isn't my suggested algorithm for deblocking interlaced video. Simply a demonstration of why you need to deinteralce before deblocking. A simple Bob() results fuzzy video with jaggies, not suitable for production. But it's fast and easy for this demonstration.

[sanlyn]

Right, I understand. I already had the pics, it's just there for comparison (how to do it the wrong way!). I'll add a note to the post.

[sanlyn]

F347-original-interlaced (below, unmodified capture from the source mpg):
There are obvious macroblocks on the big guy's bare leg, near the knee, and one at chest level.

If you Bob() you'll see that both fields are full of macroblocks. You'll also see that all the even fields have screwed up chroma (a mix of the field before and the field after).

Yes, I noticed the chroma junk after you pointed it out. Even and Odd also look like video from two different players.

BTW, I'm removing the double-Bob sample. There are other Bob pics, it's just taking up bandwidth, and many readers might not get the point.