DCTune Custom Matrix Discussion (thread-split) [Archive]

SoonUDie

26th April 2004, 15:56

It would be awesome if the process could be automated, but I don't really have a clue how to do it at the moment. It's already automated for the most part (but in different steps).

My chain of processing goes like this:
1. Use Avisynth to create frames in native DCTune format (.ppm).

Example script:

SelectEvery(last, 1000, 1)
ConvertToRGB()
ImageWriter(last, file = "C:\dctune_raw\yoursource", start = 0, end = 0, type = "ppm")

This will output every thousandth frame in the form of yoursource000000.ppm, yoursource000001.ppm, etc.

2. Use BathEncoder to call DCTune on all of the images. Simply drag and drop the files into it, set your encoding switches, and hit "go". BathEncoder creates a batchfile which calls DCTune on every image. You can get it here (http://members.home.nl/w.speek/batchenc.htm). Be sure to put DCTune (http://vision.arc.nasa.gov/dctune/) in the same folder.

The switches I normally use for DCTune are:
dctune2.0 -f <infile> -p 1 -inches 8 -dpi 133 -o <outfile>_quants_p1.txt -j <outfile>_p1.jpg

Note the -p value. This is basically your quality setting. -p 1 is normally perceptually lossless. For a high-bitrate encode, I normally use this to create my intra-matrix; then I run it a second time with -p 2 and change the outputs accordingly, and use those to create my inter-matrix.

In general, I use the following p-values:
High bitrate: 1, 2
Medium bitrat: 2, 3
Low bitrate: 3, 4

The high bitrate matrices also look good at low bitrate, but you have a higher average quantizer (so it looks a little ragged).

3. Average the results to get a good quantization matrix. I'm currently using a pretty simplistic program I wrote in Java to do this. I'm not really a programmer, so I have no clue how to export it at the moment. I *might* be able to turn it into an applet.

dimzon

26th April 2004, 16:20

Originally posted by SoonUDie
It would be awesome if the process could be automated
Seems like I can write simple avs2qmatrix.bat to automate it :)
===================
Done!
Create such directory structure:

c:\dctune\
c:\dctune\frames\
c:\dctune\trash\
c:\dctune\inter\
c:\dctune\intra\

place avs2avi.exe and dctune2.0.exe into c:\dctune\ folder

place extract.bat into c:\dctune\ folder

@echo off
del /Q c:\dctune\frames\*.ppm > NUL
echo Import("%1") > temp.avs
echo SelectEvery(last, %2, %2/2) >> temp.avs
echo ConvertToRGB24() >> temp.avs
echo ImageWriter(last, file="c:\dctune\frames\x", start=0, end=0, type="ppm") >> temp.avs
echo crop(0,0,8,8) >> temp.avs
c:\dctune\avs2avi temp.avs -o n -c DIV3

place calculate.bat into c:\dctune\ folder

@echo off
del /Q c:\dctune\intra\*.* > NUL
del /Q c:\dctune\inter\*.* > NUL
FOR %%i IN (c:\dctune\frames\*.ppm) DO c:\dctune\dctune2.0.exe -f %%i -p %1 -inches 8 -dpi 133 -o c:\dctune\intra\%%~ni.txt -j c:\dctune\trash\last.jpg
FOR %%i IN (c:\dctune\frames\*.ppm) DO c:\dctune\dctune2.0.exe -f %%i -p %2 -inches 8 -dpi 133 -o c:\dctune\inter\%%~ni.txt -j c:\dctune\trash\last.jpg
cscript c:\dctune\calculate.vbs

place calculate.vbs into c:\dctune\ folder

Option Explicit
dim g_oFS ' FileSystem Object
set g_oFS = CreateObject("Scripting.FileSystemObject")

function ReadMatrixFromFile(sFileName)
dim aResult
dim aLine
dim oFile
dim s
dim i,j
redim aResult(63) ' 8X8, zero based
set oFile = g_oFS.OpenTextFile(sFileName, 1)
' Skip file header ;)
oFile.ReadLine
oFile.ReadLine
oFile.ReadLine
oFile.ReadLine
oFile.ReadLine
oFile.ReadLine
oFile.ReadLine
oFile.ReadLine
oFile.ReadLine
oFile.ReadLine

for i=0 to 7 ' 8 lines total
s = oFile.ReadLine
s = replace( trim(s), CHR(32) & CHR(32), CHR(32))
s = replace( trim(s), CHR(32) & CHR(32), CHR(32))
s = replace( trim(s), CHR(32) & CHR(32), CHR(32))
s = replace( trim(s), CHR(32) & CHR(32), CHR(32))
s = replace( trim(s), CHR(32) & CHR(32), CHR(32))
s = replace( trim(s), CHR(32) & CHR(32), CHR(32))
aLine = Split(s,CHR(32))
for j=0 to 7
aResult(i*8+j)=CLng(aLine(j))
next
next
oFile.Close
ReadMatrixFromFile = aResult
end function

function CalculateAverageMatrix(sDirectory)
dim oFile
dim aResult
dim aSubMatrix
dim nCount
dim i
redim aResult(63) ' 8X8, zero based
for i=0 to 63
aResult(i)=CDbl(0)
next
nCount = 0
for each oFile in g_oFS.getFolder(sDirectory).files
aSubMatrix = ReadMatrixFromFile(oFile.Path)
nCount = nCount + 1
for i=0 to 63
aResult(i)=aResult(i) + CDbl(aSubMatrix(i))
next
next
if nCount>0 then
for i=0 to 63
aResult(i)=CDbl( aResult(i) /nCount)
if aResult(i) > 255 then aResult(i)=255
next
end if
CalculateAverageMatrix = aResult
end function

sub main
dim aIntraMatrix
dim aInterMatrix
dim aTotal
dim rMultiplier
dim oXml
dim oStream
dim i
redim aTotal(127)
aIntraMatrix = CalculateAverageMatrix("c:\dctune\intra")
aInterMatrix = CalculateAverageMatrix("c:\dctune\inter")
for i=0 to 63
aTotal(i)=aIntraMatrix(i)
next
for i=0 to 63
aTotal(i+64)=aInterMatrix(i)
next

'Note: The top left value (0, 0) must always be 8
' (may be defined so in the MPEG standard), and other values shall not be below.
rMultiplier = 8/aTotal(0)
for i=0 to 63
aTotal(i)=CLng( rMultiplier * CDbl(aTotal(i)))
if aTotal(i) > 255 then
aTotal(i) = 255
elseif aTotal(i) < 8 then
aTotal(i) = 8
end if
next

'Note: The top left value (0, 0) must always be 16
' (may be defined so in the MPEG standard), and other values shall not be below.
rMultiplier = 16/aTotal(64)
for i=64 to 127
aTotal(i)=CLng( rMultiplier * CDbl(aTotal(i)))
if aTotal(i) > 255 then
aTotal(i) = 255
elseif aTotal(i) < 16 then
aTotal(i) = 16
end if
next

for i=0 to 127
if aTotal(i)>15 then
aTotal(i) = HEX(aTotal(i))
else
aTotal(i) = "0" & HEX(aTotal(i))
end if
next
set oXml = CreateObject("MSXML2.DOMDocument.3.0").createElement("dummy")
oXml.dataType="bin.hex"
oXml.text=join(aTotal,"")
set oStream = CreateObject("ADODB.Stream")
oStream.Open
oStream.Type=1
oStream.Write oXml.nodeTypedValue
oStream.SaveToFile "c:\dctune\result.qmatrix", 2
oStream.Close
MsgBox "Done!"
end sub

CALL Main()

USAGE

1-st - extract frames
extract.bat <source_avs_file_name_and_path> <frame_denominator>
example: extract c:\dvdrip\gigli\gigli.avs 1000

2-nd - calculate qmatrix
calculate.bat <intra_quality>, <inter_quality>
example: calculate 2 3

3-td - got Your matrix c:\dctune\result.qmatrix

SoonUDie

26th April 2004, 18:59

I bow before your skills. :goodpost:

Another note: you should probably go through and manually check all the frames, and delete any ones that you don't want to use. You should definately delete blank frames (all black or all white), and you may want to cut out any credits as well.

vaxis

26th April 2004, 20:22

This is really cool. I am in the middle of testing it on Iron Maiden - Rock In Rio, which is the hardiest dvd to encode I have ever seen.

I did have to change a couple of lines in the calculate.vbs to get it to work. In the function ReadMatrixFromFile, where the double for loop is, your way of dividing the numbers into the array aLine did not work, some post in it where just blanks and the CLng then returned an error. instead I simply made aLine to a string and implemented the inner for loop a little different and removed the lines with replace calls in the outer for loop.

So here is the loop solution that worked for me:

for i=0 to 7 ' 8 lines total
s = oFile.ReadLine
for j=0 to 7
aLine = trim(mid(s,j*4+1,4))
aResult(i*8+j)=CLng(aLine)
next
next

You should update your script with that.

Again, thanks for this much easier automatic solution.

SoonUDie

27th April 2004, 00:41

I hope it works out for you. In general, I've found that it seems to produce a very good amount of detail - more than HVS Best is most cases. The main plus side, in my mind, is that the matrix is totally adapted to the source. I think the only way you could improve upon this general method would be if XviD automatically calculated and used a different (optimized) quant matrix for each individual scene.

SoonUDie

27th April 2004, 00:45

Oh, I should mention that DCTune produces matrices with a lowest possible value of 2. You may want to implement a way to multiply all the other values accordingly, so that you always get 8 in the upperp left corner. If not, you can use excel and a little bit of math once you have the final matrix.

ObiKenobi

27th April 2004, 01:26

Just to make sure I'm doing this right, when DCTune outputs the txt file like the one below, I use the top matrix right?

# Quantization Matrices for G:\000000.PPM
# Using DCTune Version 2.0
# Quantization Table Calculated Using Perceptual Optimization
# Target Perceptual Error 2.000000
# Mean Luminance: 33.532667 Candelas/meter^2
# Peak Luminance: 67.162400 Candelas/meter^2
# Number Gray Steps: 256
# Pixels/Degree X: 18.666667
# Pixels/Degree Y: 18.666667
# Size of DCT: 8
4 4 4 4 4 6 255 255
4 2 4 2 4 4 255 255
4 4 6 4 4 6 255 255
4 4 4 6 10 255 255 255
4 4 4 10 255 255 255 255
6 4 255 255 255 255 255 255
8 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255
# Quantization Table Calculated Using Perceptual Optimization
# Target Perceptual Error 2.000000
# Mean Luminance: 33.532667 Candelas/meter^2
# Peak Luminance: 67.162400 Candelas/meter^2
# Number Gray Steps: 256
# Pixels/Degree X: 9.333333
# Pixels/Degree Y: 9.333333
# Size of DCT: 8
18 255 255 255 255 255 255 255
24 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255
# Quantization Table Calculated Using Perceptual Optimization
# Target Perceptual Error 2.000000
# Mean Luminance: 33.532667 Candelas/meter^2
# Peak Luminance: 67.162400 Candelas/meter^2
# Number Gray Steps: 256
# Pixels/Degree X: 9.333333
# Pixels/Degree Y: 9.333333
# Size of DCT: 8
8 255 255 255 255 255 255 255
8 255 255 255 255 255 255 255
8 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255

iradic

27th April 2004, 02:10

hi...

for me -o c:\dctune\inter\%%~ni.txt doesnt work... only -o %%i.txt
using winme

any help...

EDIT: problem solved

SoonUDie

27th April 2004, 02:53

Originally posted by ObiKenobi
Just to make sure I'm doing this right, when DCTune outputs the txt file like the one below, I use the top matrix right?Yes. I believe the bottom ones are for chroma (not an option with XviD... and even if it was, DCTune tends to suck at chroma except at -p1 or higher quality)

ObiKenobi

27th April 2004, 02:55

Is it just me or does that matrix seem a bit odd though? Compared to other matrices I've looked at it just looks so much different.

SoonUDie

27th April 2004, 08:10

Originally posted by ObiKenobi
Is it just me or does that matrix seem a bit odd though? Compared to other matrices I've looked at it just looks so much different. If I understand quant matrices correctly, each entry covers a specified frequency range. Since humans don't perceive these frequencies equally, an optimized matrix will increase the quantization of those frequencies. I believe that the MPEG quant works this way (but uses an older percptual model?), and I know that the HVS matrices (HVS = Human Visual System) work that way, too.

The difference between using a generalized matrix like HVS-best and calculating an average matrix using DCTune is that, theoretically, the averaging method will create a perceptual matrix adapted to the source's quality, and therefore better for purposes of compression, retaining the same amount of quality with less bits. Of course, this assumes that DCTune's alogrithms for creating an optimized matrix are valid.

As an example of the DCTune method, here is the average matrix I got for the SeaBisucit R1 (widescreen) DVD at p=1 (~400 frames sampled, using my program to calculate the matrix):
8 8 8 10 11 16 86 255
8 8 9 9 10 28 106 255
8 9 16 21 35 77 255 255
8 9 20 60 87 174 255 255
10 10 23 86 149 255 255 255
15 16 57 118 255 255 255 255
26 22 92 185 255 255 255 255
41 59 166 255 255 255 255 255

dimzon

27th April 2004, 09:31

Originally posted by vaxis
I did have to change a couple of lines in the calculate.vbs to get it to work. In the function ReadMatrixFromFile, where the double for loop is, your way of dividing the numbers into the array aLine did not work, some post in it where just blanks and the CLng then returned an error. instead I simply made aLine to a string and implemented the inner for loop a little different and removed the lines with replace calls in the outer for loop.

It's bcz forum engine changed my source string:

s = replace( trim(s), "<SPACE><SPACE>", "<SPACE>")

to

s = replace( trim(s), "<SPACE>", "<SPACE>")

so i will change this string into

s = replace( trim(s), CHR(32) & CHR(32), CHR(32))

:)

anywhere Your solution works fine too (but i'm too lazy to analyze quantizers position into source string)

and some user nickname=Shade PM-ed me such info about DCTune (i have no idea why he PM-ed to me but not to post this to this thread so I repost it):

DCTune can only calculate intra matrix. It's designed for still image(JPEG), not for motion pictures.
The second matrix generated by DCTune is for chroma. JPEG can use different matrix for luma and chroma, and both matrix are intra matrix. You must design inter matrix(for inter block in P/B-frame) by youself.
Intra matrix quantize intra-block, the coefficients in intra-block are "texture".
Inter matrix quantize inter-block, the coefficients in inter-block are "residual data after ME".
The properties of intra/inter matrix are different.
:)

SoonUDie

27th April 2004, 10:42

"DCTune can only calculate intra matrix. It's designed for still image(JPEG), not for motion pictures."True. I don't really know as much about quantization as I want to, yet, but I've just assumed that inter-frames in general use a "lower-quality" matrix (which is where my use of 1 lower p value comes from).

Guys, can you remember me? I'm the one who started this thread :DYea, sorry about that :D. I've somewhat hijacked the thread in developing my own method for creating matrices.

Also, you can stop using the Evil Matrix now - it's served its purpose as a test, and I think my current method far outstripes it in creating a quality encode. I would appreciate it if you try my new method :)

On another point, I went through the forums using search, and DCTune has been discussed in relation to XviD a few times before (though not in this capacity or detail), and some people came to the same conclusion I have - it would be awesome if we could use use DCTune to optimize individual scenes, or even frames. For example: call DCTune on each i-frame, so that each scene gets the most appropriate matrix. I know that XviD supports (or at least, used to support) changing quantizations like this, but I think it's not mpeg4 compliant.

dimzon

27th April 2004, 10:58

Originally posted by SoonUDie
Oh, I should mention that DCTune produces matrices with a lowest possible value of 2. You may want to implement a way to multiply all the other values accordingly, so that you always get 8 in the upperp left corner. If not, you can use excel and a little bit of math once you have the final matrix.
How does You thik what is the best:
1) multiply all the other values on (8/top_left_intra_value)
OR
2) increase all the other values on (8-top_left_intra_value)
OR
3) set values=8 where values is < 8

dimzon

27th April 2004, 12:34

I have just updated calculate.vbs in my post to follow MPEG specs.

That's my quantizer matrix for "KNOCKIN' ON HEAVEN'S DOOR" movie

8,8,8,8,11,16,30,114
8,8,8,8,10,14,25,73
8,8,13,14,20,26,45,151
8,8,14,24,43,63,123,255
12,10,19,38,114,200,255,255
16,12,25,56,208,255,255,255
27,23,44,125,255,255,255,255
85,61,140,255,255,255,255,255

16,16,16,17,29,53,152,255
16,16,18,17,29,48,128,255
16,18,32,39,61,108,255,255
17,17,39,80,181,255,255,255
30,28,57,178,255,255,255,255
48,42,100,255,255,255,255,255
143,117,255,255,255,255,255,255
255,255,255,255,255,255,255,255

now I'm encoding...

springl

27th April 2004, 14:08

@dimzon
How do I have to modify the bat files to make them work flawlessly under WIN98SE,'cause I get a "not valid option /Q" when I run extract.bat and calculate.bat. Moreover,due to calculate.vbs, I get
a runtime error of M$ Vbscript : "activeX is not able to create the
object 'MSXML2.DOMDocument.3.0'".And of course, result.qmatrix is not created.

Thanks.

dimzon

27th April 2004, 14:17

@springl
you need Microsoft XML Parser 3.0 SP4 to run calculate.vbs
get it here:
http://msdn.microsoft.com/XML/XMLDownloads/default.aspx

about bat files you can try such workaround:

1) create such yes.txt into c:\dctune\ folder:
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y

2) REPLACE
DEL /Q {something} >NUL
WITH
DEL {something} < c:\dctune\yes.txt >NUL

Good luck

springl

27th April 2004, 15:07

The old problems are gone thanks to XML Parser and your workaround but now there's a new runtime error of M$ VBScript :activeX is not able to create object 'ADODB.stream'.
Result.qmatrix is not created.:confused:

dimzon

27th April 2004, 15:26

Originally posted by springl
new runtime error of M$ VBScript :activeX is not able to create object 'ADODB.stream'.
You need Microsoft Data Access Components to. Sorry, it is my mistake, download it here:
http://www.microsoft.com/downloads/details.aspx?familyid=9ad000f2-cae7-493d-b0f3-ae36c570ade8&languageid=f49e8428-7071-4979-8a67-3cffcb0c2524&displaylang=en

Explanations
I used MSXML Parser to convert matrix data into raw bytearray (VBScript can't work with bytearray directly so there are some programming trick )
I used ADODB.Stream (part of MDAC) to store raw bytearray as result.qmatrix (VBScript can't work with files directly and Scripting.FileSystemObject can't work with binary files to so there are some another programming trick)

springl

27th April 2004, 16:09

@dimzon
OK,finally I've got the matrix.
Thank you very much for your help.

:)

SoonUDie

27th April 2004, 16:20

Hey guys - I've something interesting in my testing. For 1-CD encodes, a p 2/3 or p 3/4 mix works well... but I seem to be getting even better results by averaging a p 1 matrix with HVS-best. This version isn't a clear winner in detail, but it absolutely kills artifacts, and in general the "structure" of the image is preserved better.

dimzon

27th April 2004, 16:47

Originally posted by SoonUDie
Hey guys - I've something interesting in my testing. For 1-CD encodes, a p 2/3 or p 3/4 mix works well... but I seem to be getting even better results by averaging a p 1 matrix with HVS-best. This version isn't a clear winner in detail, but it absolutely kills artifacts, and in general the "structure" of the image is preserved better.

I have updated calculate.vbs so it produce now 2 files - result.qmatrix and result-mix-with-hvs-best.qmatrix

calculate.vbs

Option Explicit
dim g_oFS ' FileSystem Object
set g_oFS = CreateObject("Scripting.FileSystemObject")

function ReadMatrixFromFile(sFileName)
dim aResult
dim aLine
dim oFile
dim s
dim i,j
redim aResult(63) ' 8X8, zero based
set oFile = g_oFS.OpenTextFile(sFileName, 1)
' Skip file header ;)
oFile.ReadLine
oFile.ReadLine
oFile.ReadLine
oFile.ReadLine
oFile.ReadLine
oFile.ReadLine
oFile.ReadLine
oFile.ReadLine
oFile.ReadLine
oFile.ReadLine

for i=0 to 7 ' 8 lines total
s = oFile.ReadLine
s = replace( trim(s), CHR(32) & CHR(32), CHR(32))
s = replace( trim(s), CHR(32) & CHR(32), CHR(32))
s = replace( trim(s), CHR(32) & CHR(32), CHR(32))
s = replace( trim(s), CHR(32) & CHR(32), CHR(32))
s = replace( trim(s), CHR(32) & CHR(32), CHR(32))
s = replace( trim(s), CHR(32) & CHR(32), CHR(32))
aLine = Split(s,CHR(32))
for j=0 to 7
aResult(i*8+j)=CLng(aLine(j))
next
next
oFile.Close
ReadMatrixFromFile = aResult
end function

function CalculateAverageMatrix(sDirectory)
dim oFile
dim aResult
dim aSubMatrix
dim nCount
dim i
redim aResult(63) ' 8X8, zero based
for i=0 to 63
aResult(i)=CDbl(0)
next
nCount = 0
for each oFile in g_oFS.getFolder(sDirectory).files
aSubMatrix = ReadMatrixFromFile(oFile.Path)
nCount = nCount + 1
for i=0 to 63
aResult(i)=aResult(i) + CDbl(aSubMatrix(i))
next
next
if nCount>0 then
for i=0 to 63
aResult(i)=CDbl( aResult(i) /nCount)
if aResult(i) > 255 then aResult(i)=255
next
end if
CalculateAverageMatrix = aResult
end function

Sub SaveMatriX(aMatrix, sFileName)
dim aTotal
dim oXml
dim oStream
dim i
redim aTotal(127)
for i=0 to 127
if aMatrix(i)>15 then
aTotal(i) = HEX(aMatrix(i))
else
aTotal(i) = "0" & HEX(aMatrix(i))
end if
next
set oXml = CreateObject("MSXML2.DOMDocument.3.0").createElement("dummy")
oXml.dataType="bin.hex"
oXml.text=join(aTotal,"")
set oStream = CreateObject("ADODB.Stream")
oStream.Open
oStream.Type=1
oStream.Write oXml.nodeTypedValue
oStream.SaveToFile sFileName, 2
oStream.Close
End Sub

function NormalizeMatrix(aMatrix)
dim aTotal, i
dim rMultiplier
aTotal = Split(Join(aMatrix,"???") ,"???")
'Note: The top left value (0, 0) must always be 8
' (may be defined so in the MPEG standard), and other values shall not be below.
rMultiplier = 8/aTotal(0)
for i=0 to 63
aTotal(i)=CLng( rMultiplier * CDbl(aTotal(i)))
if aTotal(i) > 255 then
aTotal(i) = 255
elseif aTotal(i) < 8 then
aTotal(i) = 8
end if
next

'Note: The top left value (0, 0) must always be 16
' (may be defined so in the MPEG standard), and other values shall not be below.
rMultiplier = 16/aTotal(64)
for i=64 to 127
aTotal(i)=CLng( rMultiplier * CDbl(aTotal(i)))
if aTotal(i) > 255 then
aTotal(i) = 255
elseif aTotal(i) < 16 then
aTotal(i) = 16
end if
next
NormalizeMatrix = aTotal
end function

sub main
const HVS_BEST="8,16,16,16,17,18,21,24,16,16,16,16,17,19,22,25,16,16,17,18,20,22,25,29,16,16,18,21,24,27,31,36,17,17,20,24,30,35,41,47,18,19,22,27,35,44,54,65,21,22,25,31,41,54,70,88,24,25,29,36,47,65,88,115,18,18,18,18,19,21,23,27,18,18,18,18,19,21,24,29,18,18,19,20,22,24,28,32,18,18,20,24,27,30,35,40,19,19,22,27,33,39,46,53,21,21,24,30,39,50,61,73,23,24,28,35,46,61,79,98,27,29,32,40,53,73,98,129"
dim aIntraMatrix
dim aInterMatrix
dim aTotal
dim i
redim aTotal(127)
aIntraMatrix = CalculateAverageMatrix("c:\dctune\intra")
aInterMatrix = CalculateAverageMatrix("c:\dctune\inter")
for i=0 to 63
aTotal(i)=aIntraMatrix(i)
next
for i=0 to 63
aTotal(i+64)=aInterMatrix(i)
next
SaveMatriX NormalizeMatrix(aTotal), "c:\dctune\result.qmatrix"
' Mix with hvs-best
aIntraMatrix = Split(HVS_BEST, ",")
for i=0 to 127
aTotal(i) = (aTotal(i)+aIntraMatrix(i))/2
next
SaveMatriX NormalizeMatrix(aTotal), "c:\dctune\result-mix-with-hvs-best.qmatrix"
MsgBox "Done!"
end sub

CALL Main()

dimzon

28th April 2004, 10:49

"KNOCKIN' ON HEAVEN'S DOOR" movie
http://img10.imageshack.us/img10/3295/frm-000114.png
http://img10.imageshack.us/img10/958/frm-000006.png
most of my frames is more than 600k :(
So I converted them to JPEG @ Q=100%
http://img9.imageshack.us/img9/3395/frm-000088.JPG
http://img7.imageshack.us/img7/1536/frm-000021.JPG
http://img9.imageshack.us/img9/4343/frm-000044.JPG
http://img7.imageshack.us/img7/6242/frm-000169.JPG
http://img10.imageshack.us/img10/1310/frm-000018.JPG
http://img7.imageshack.us/img7/6890/frm-000102.JPG
http://img8.imageshack.us/img8/308/frm-000152.JPG

SoonUDie

28th April 2004, 12:11

Originally posted by dimzon
most of my frames is more than 600k :( In general, jpeg with a very high quality setting will look almost visually identical even at close inspection. PNG is unnecessary unless the differences between frames are extremely close.

dimzon

29th April 2004, 10:41

@All
I can't see difference between Jawor matrix and DCTune matrix (but quantizer values are different):

it's DCTune(1,2)

8,8,8,8,11,16,30,114
8,8,8,8,10,14,25,73
8,8,13,14,20,26,45,151
8,8,14,24,43,63,123,255
12,10,19,38,114,200,255,255
16,12,25,56,208,255,255,255
27,23,44,125,255,255,255,255
85,61,140,255,255,255,255,255

16,16,16,17,29,53,152,255
16,16,18,17,29,48,128,255
16,18,32,39,61,108,255,255
17,17,39,80,181,255,255,255
30,28,57,178,255,255,255,255
48,42,100,255,255,255,255,255
143,117,255,255,255,255,255,255
255,255,255,255,255,255,255,255

It's DCTune(1,2) averaged with HVS-Best-Picture

8,14,14,14,15,18,23,48
14,13,14,14,15,18,23,38
14,14,16,17,20,23,30,61
14,14,17,22,29,36,55,120
16,15,20,28,52,79,135,197
17,17,23,35,81,130,188,231
23,22,30,56,135,190,231,252
40,35,58,114,192,227,251,255

16,16,16,16,20,27,53,138
16,16,16,16,20,26,48,131
16,16,21,23,30,43,84,164
16,16,23,36,62,93,148,193
20,20,29,62,133,166,195,206
26,24,41,95,167,196,212,220
51,45,85,148,195,211,224,237
119,115,157,190,206,220,237,255

It's source.avs

Mpeg2Source("ddn.d2v")
trim(20*25+1,0)
CROP(8,58,-8,-58)
Tweak(sat=1.5, coring=false)
Undot()
LanczosResize(640,384)
removeDirt(show=0)
Undot()
Deen()
VagueDenoiser(threshold=0.8,method=1,nsteps=6,chroma=true)

It's compare.avs

vOriginal =import("c:\dctune\1\source.avs").Subtitle("ORIGINAL").trim(0,-100000)
vHvsGood = aviSource("c:\dctune\1\d640.avi").Subtitle("Jawor_1CD_Matrix").trim(1,-100000)
vDCTune12 = aviSource("c:\dctune\1\d_dctune_12.avi").Subtitle("DCTune(1,2)").trim(1,-100000)
vDCTune12Mix = aviSource("c:\dctune\1\d_dctune_12_mix_hvs_best.avi").Subtitle("DCTune(1,2)+HvsBest").trim(1,-100000)
StackVertical(vOriginal,vHvsGood,vDCTune12,vDCTune12Mix)
SelectEvery(last, 500, 250)
ConvertToRGB24()
ImageWriter(last, file="c:\dctune\1\frm-", start=0, end=0, type="png")
crop(0,0,8,8)

May be anyone can point me on visual diff?

there are frames (jpeg q=100%)
http://img7.imageshack.us/img7/8044/frm-000190.JPG
http://img8.imageshack.us/img8/6242/frm-000169.JPG
http://img9.imageshack.us/img9/4816/frm-000162.JPG
http://img7.imageshack.us/img7/3923/frm-103.JPG
http://img8.imageshack.us/img8/4284/frm-000082.JPG

Maybe I need encode at higher resolution?
it's "KNOCKIN' ON HEAVEN'S DOOR" movie 1hr28m 640x384@1000kbps encode

SoonUDie

29th April 2004, 12:28

How exactly did you combine HVS with your matrices? I was assuming you'd use a standard average, producing a matrix like this (from your numbers):8,12,12,12,14,17,26,69
12,12,12,12,14,17,24,49
12,12,15,16,20,24,35,90
12,12,16,23,34,45,77,146
15,14,20,31,72,118,148,151
17,16,24,42,122,150,155,160
24,23,35,78,148,155,163,172
55,43,85,146,151,160,172,185

17,17,17,18,24,37,88,141
17,17,18,18,24,35,76,142
17,18,26,30,42,66,142,144
18,18,30,52,104,143,145,148
25,24,40,103,144,147,151,154
35,32,62,143,147,153,158,164
83,71,142,145,151,158,167,177
141,142,144,148,154,164,177,192
Which you can see is very different...

Also, when you do your conversion to make the DCTune matrix have 8 in the upper left corner, do you use adding or multiplying? I've been using multiplication.

dimzon

29th April 2004, 12:35

for i=0 to 127
aTotal(i) = (aTotal(i)+aHvsBestatrix(i))/2
next

'Note: The top left value (0, 0) must always be 8
' (may be defined so in the MPEG standard), and other values shall not be below.
rMultiplier = 8/aTotal(0)
for i=0 to 63
aTotal(i)=CLng( rMultiplier * CDbl(aTotal(i)))
if aTotal(i) > 255 then
aTotal(i) = 255
elseif aTotal(i) < 8 then
aTotal(i) = 8
end if
next

'Note: The top left value (0, 0) must always be 16
' (may be defined so in the MPEG standard), and other values shall not be below.
rMultiplier = 16/aTotal(64)
for i=64 to 127
aTotal(i)=CLng( rMultiplier * CDbl(aTotal(i)))
if aTotal(i) > 255 then
aTotal(i) = 255
elseif aTotal(i) < 16 then
aTotal(i) = 16
end if
next

maybe i have incorrect hvs-best-matrix?

8,16,16,16,17,18,21,24
16,16,16,16,17,19,22,25
16,16,17,18,20,22,25,29
16,16,18,21,24,27,31,36
17,17,20,24,30,35,41,47
18,19,22,27,35,44,54,65
21,22,25,31,41,54,70,88
24,25,29,36,47,65,88,115

18,18,18,18,19,21,23,27
18,18,18,18,19,21,24,29
18,18,19,20,22,24,28,32
18,18,20,24,27,30,35,40
19,19,22,27,33,39,46,53
21,21,24,30,39,50,61,73
23,24,28,35,46,61,79,98
27,29,32,40,53,73,98,129

Oh, I know!

My algorythm is

OBTAIN_AVERAGE_DCTUNE
AVERAGE_IT_WITH_HVS_BEST
NORMALIZE_IT

Your algorythm is

OBTAIN_AVERAGE_DCTUNE
NORMALIZE_IT
AVERAGE_IT_WITH_HVS_BEST

Colyptic

29th April 2004, 14:01

Interesting about the DCTune method, I'm one to believe that there is no perfect matrix for every movie out there and this may be a step towards finding the right matrix or at least right kind of matrix for each movie.

However, I ran into an error with calculate.vbs gives me 'Microsoft VBScript error: division by zero line:99 column:1

which I believe is this line:
rMultiplier = 8/aTotal(0)

any help would be appreciated.

dimzon

29th April 2004, 14:05

Originally posted by Colyptic

However, I ran into an error with calculate.vbs gives me 'Microsoft VBScript error: division by zero line:99 column:1

which I believe is this line:
rMultiplier = 8/aTotal(0)

any help would be appreciated.

I think you forgot generate quantizer matrixes by DCTune (c:\dctune\intra or/and c:\dctune\inter is empty)

Read full process cafely (bottom link in my signature)

Colyptic

29th April 2004, 14:13

wow that was a quick reply, I figured out what I was doing wrong. Wasn't entering an integer for extract.bat so avs2avi didn't output any .ppm frames. I was testing with a 15 second clip and assumed leaving out the frame interval would encode the whole thing. But need to use at least 1 there.

dimzon

29th April 2004, 14:15

Originally posted by Colyptic
But need to use at least 1 there.
suggest at least 2.

bill_baroud

29th April 2004, 15:44

interesting thread about DCTune... i'm reading the papers atm, but i didn't find any source code, too bad :/

now i am dreaming of a new first pass in xvid outputting the "best" matrix for the movie :D

SoonUDie

29th April 2004, 15:56

Originally posted by dimzon
Your algorythm is

OBTAIN_AVERAGE_DCTUNE
NORMALIZE_IT
AVERAGE_IT_WITH_HVS_BEST
Correct, because HVS-Best is already "normalized", and if you averaged them before normalizing the DCTune matrix, you wouldn't be "weighing" them equally. You can maintain precision by not rounding off the values until you have the final blended matrix :)

About the 16 in the interframes: I don't know what values are mpeg4 and standalone compliant, but the XviD decoder hasn't had any problems decoding non-16 matrices.

DarkWave

29th April 2004, 16:51

@dimzon & SoonUDie

is it right then to change dimzon´s script like this:
sub main
...
SaveMatriX NormalizeMatrix(aTotal), "c:\dctune\result.qmatrix"
' Mix with hvs-best
aIntraMatrix = Split(HVS_BEST, ",")
aTotal = NormalizeMatrix(aTotal)
for i=0 to 127
aTotal(i) = (aTotal(i)+aIntraMatrix(i))/2
next
SaveMatriX aTotal, "c:\dctune\result-mix-with-hvs-best.qmatrix"
MsgBox "Done!"
end sub

greetz
DarkWave

dimzon

29th April 2004, 16:57

Originally posted by DarkWave
@dimzon & SoonUDie

is it right then to change dimzon´s script like this:
sub main
...
SaveMatriX NormalizeMatrix(aTotal), "c:\dctune\result.qmatrix"
' Mix with hvs-best
aIntraMatrix = Split(HVS_BEST, ",")
aTotal = NormalizeMatrix(aTotal)
for i=0 to 127
aTotal(i) = (aTotal(i)+aIntraMatrix(i))/2
next
SaveMatriX aTotal, "c:\dctune\result-mix-with-hvs-best.qmatrix"
MsgBox "Done!"
end sub

greetz
DarkWave
correct

echo-kk

29th April 2004, 23:43

@dimzon
What would you consider a good 'frame_denominator' value for the resulting matrice to be representative of the movie?

I'm asking because yesterday I encoded a short 3.5min test clip with the following matrices: MPEG, HVS-Best, DCTune(1,1), DCTune(1,2).
The DCTune matrices were calculated for the full movie (without credits) with a frame denominator of 250 (every 10sec - PAL source).

My results: Although the DCTune matrices also did a good job I preferred the MPEG/HVS-Best test clips.
Maybe a frame denominator of 250 simply was to high for this movie (a mixture of quite a lot of still scenes and action mostly taking place in smoke/fog)?

Right now I have set up a new test for a different movie. I'm calculating the matrices with a frame denominator of 50 (every 2sec!) and am planing to encode longer test clips.
I sure hope this test will yield a different result (which I'll post when ready)!

SoonUDie

30th April 2004, 02:57

@echo-kk
Have you tried mixing DCTune(1,1) with HVS-Best? I have found that this seems to produce a high-detailed result with little to no artifacts; in fact it seems to look better than either HVS-Best or DCTune by themselves.

Also, I do not believe DCTune will benefit significantly from increasing the number of shots so much, but I could be wrong... I just don't have the will power to wait for the creation and analysis of >3000 frames :eek:

kilg0r3

30th April 2004, 07:30

The young and wild at work. go Go GO!

CruNcher

30th April 2004, 09:52

The young and wild at work. go Go GO!

yeah but the question is is this scripting marathon worth the effort without one of the Nasa guys Help or the source of dctune somehow i doubt it.

SoonUDie

30th April 2004, 10:09

I have found it useful personally - in 90% of cases where I've tested it, a DCTune(1,1)/HVS matrix will beat out all others at medium to low bitrates.

Tweaking a quantization matrix only provides a small amount of improvement in quality compared to the other steps in making a good encode (filters, other encoding settings, etc.), but maximum quality is my goal, so naturally trying new methods to find better quantizers is important to me :)

echo-kk

1st May 2004, 13:12

Here are the results of my 2nd test with DCTune...

Setup:
704x432 (anamorphic); no Filters; 1000kbps -> 0.138 Bits/(Pixel*Frame).
XviD: Defaults with Adapt. Quant, no packed bitstream, VHQ 4, Chroma Motion & optimizer, 250 max I-frame, Trellis.
Resulting test clips: length 5min (-> 'SelectRangeEvery' with random 10sec long scenes).

DCTune matrices were calculated for the whole movie (except credits) with frame denominators of 50 and 1500. HVS_old refers to dimzon's standard DCTune/HVS-mix calculation; HVS_new to DarkWave's/SoonUDie's proposed script change.

I normally only trust my eyes, so I was quite astonished that the different quality metrics seemed to confirm my (subjective) visual impressions...
avg. overall avg.
SSIM PSNR VQM
MPEG 84.69 46.7491 0.459957
HVS-Best 84.76 46.7466 0.461762
DCTune(1,1)/HVS_old;1500 84.70 46.7375 0.460504
DCTune(1,1)/HVS_old;50 84.69 46.7293 0.461257
DCTune(1,1)/HVS_new;50 84.65 46.6949 0.461896
DCTune(1,1)/HVS_new;1500 84.63 46.6875 0.462485
DCTune(1,1);1500 84.33 46.5404 0.469354
DCTune(1,1);50 84.33 46.5370 0.469403

SSIM, PSNR -> higher is better
VQM -> lower is better
You'll notice that the 'more' DCTune in the matrice the more inferior the result (both visual and mathematical)!
To my surprise DCTune also did not benefit from increasing the number of frames at all. :confused:

In both my tests I preferred the MPEG/HVS-Best clips to the DCTune ones (although the actual perceived quality difference was very small - I had to look very close). This whole idea of generating a special matrice with DCTune sounded very promising but after testing I'll just stick to MPEG or HVS-Best for the time being! :(

@SoonUDiein 90% of cases where I've tested it, a DCTune(1,1)/HVS matrix will beat out all others at medium to low bitratesBeat out all others? What matrices did you compare DCTune(1,1)/HVS to?
I suspect the 'HVS' in your 'DCTune(1,1)/HVS' actually did the job. :D
I just don't have the will power to wait for the creation and analysis of >3000 frames3000 frames? Thats only a single frame every 2 minutes of your source! I wouldn't call a matrice generated from this small amount of data very representative of the specific movie.
Now I 'know' that the 'HVS' in your 'DCTune(1,1)/HVS' actually did the job. ;)

What a pity, this seemed so promising...

Teegedeck

1st May 2004, 14:03

Thanks for your impressive test! Discarding the DCTune concept altogether though seems premature.

echo-kk

1st May 2004, 14:40

Discarding the DCTune concept altogether though seems prematureOf course you're right! Maybe my post sounds to harsh.

I also do believe there still is room for improvement (maybe in the way the original DCTune data is 'averaged', 'weighed' or 'upscaled' to be Xvid conform etc.).

I just meant that I personally won't be using a DCTune generated matrice in its current development state.
Of course I'd be willing to test if someone comes up with some new ideas. :D

Manao

1st May 2004, 14:49

Thanks for splitting the thread, I would not have discovered this interesting topic :)

A few remarks first :
- computing an inter matrix with DCTunes is useless. It can only be used to compute an intra matrix.
- So, no need to take sequences of 10 seconds randomly in the movie, it would be better to choose only keyframes.
- dimzon : posting very big pictures is interesting ( though it takes a long time to download ), but it would be even better if you said what was the quantizer used, and if it was I / P or B frames ( I can't see any differences, that's why I ask ). It would be great also if you splitted the pictures, in order us not to have to.
- SoonUDie : don't be so optimistic towards this method. The fact that you need to merge an existing matrix with the one output by DCTunes is a rather disappointing fact.

Now, a few questions, to echo-kk : what was the movie, what was the quant distribution ? Because your results may be simply explained if the bitrate was high for the movie ( since coefficients for the HVS-Best / MPEG matrix are lower than those of the other matrices ).

echo-kk

1st May 2004, 16:52

The fact that you need to merge an existing matrix with the one output by DCTunes is a rather disappointing fact.I completely agree!

a few questions, to echo-kk : what was the movie, what was the quant distribution ? Because your results may be simply explained if the bitrate was high for the movie
Master&Commander (Pal)
480x416 (wanted to test horizontal resizing only) @ 710kbps
-> rather low bitrate, but as you can see from a previous post the results did not substantially differ from my last test, the DCTune clips were inferior!
(Cannot post quant distribution because I didn't take notes and don't have the clips anymore)

Reservoir Dogs (Pal)
704x432 @ 1025kbps -> Bits per Pixel: 0.138
I wouldn't exactly call this a 'high' bitrate for this resolution, but the movie seems to compress very well.

Quant Distribution etc.
MPEG HVS-Best DCTune(1,1)50

I Frames: 1.22% 1.22% 1.22%
P Frames: 37.84% 37.82% 37.88%
B Frames: 60.95% 60.96% 60.90%

Q1&2: 1167 17.3% 1174 17.4% 934 13.9%
Q3: 1225 18.2% 1327 19.7% 1308 19.4%
Q4: 3458 51.3% 3444 51.1% 3209 47.6%
Q5: 891 13.2% 796 11.8% 1239 18.4%
Q6: 48 0.7%
Q7: 3 ~0%

AverageKeyQ: 2.52 2.55 2.95
AverageQ: 3.59 3.56 3.72

This is what DCTune delivered after several hours :D of computing...
DCTune(1,1)50 Matrix

8 8 8 8 8 10 17 162
8 8 8 8 8 8 15 157
8 8 10 10 10 13 38 255
8 8 10 15 25 41 232 255
8 8 12 37 99 255 255 255
13 11 42 181 255 255 255 255
98 116 255 255 255 255 255 255
255 255 255 255 255 255 255 255

16 16 16 16 16 19 35 255
16 16 16 16 16 16 29 255
16 16 19 19 20 27 76 255
16 16 20 30 50 82 255 255
17 16 23 75 199 255 255 255
26 22 84 255 255 255 255 255
196 231 255 255 255 255 255 255
255 255 255 255 255 255 255 255

For me the specially computed matrices produced inferior results in both medium and low bitrate scenarios.

iradic

1st May 2004, 21:33

HI
i dont think these quant stats really show the video quality when we are talking about dctune matrices... beacouse they are optimized - so one (stupid) conclusion is that dctune matrix quants should be higher than ordinary ones - producing visually the same quality at higher quants... this again leave us to our eyes to believe what is better

well opposite opinion is - dctune quants should be lower - and yes this is also true (they can also be the same) - tired of explaining ... but i think you can figure out why... :)

there is a pdf document about dctune algo - http://vision.arc.nasa.gov/publications/icip94.pdf

bye...

SoonUDie

2nd May 2004, 02:17

Thanks for the link :)

All my quality assesments so far are subjective. From my point of view, DCTune on its own does not fare well - it generally blurs more (but has less blocking) than other matrices. However, the results when I combine a DCTune matrix with HVS are, to my eyes, better than either one on its own.

I will upload a proper screenshot comparison as soon as I can, but it's finals week here :(

Colyptic

2nd May 2004, 11:16

As for not being able to create an inter-matrix and only intra-matrix. I disagree, I think its really the other way around where you should only create inter-matrix with it. I can see why you came up with that though, with intra-frames being still picture and inter-frames being motion predicted.

I've been doing some testing around with it mixed with HVSBest and first thing I noticed was huge intra frames and was kind of disappointed with the results. After looking at the intra-matrix DCTune gave me it was easy to see why the keyframes were so big. So I decided to change the intra-matrix to the MPEG standard and turned it into a low-mid bitrate matrice (about the same compressibility as HVSBest) rather than what I'd consider a mid-high bitrate matrice (produced larger files than MPEG).

It is kind of annoying to have to change them all the time and I'd really appreciate it if someone could tell me what code to change to use the MPEG standard as the intra-frame matrix all the time and take out the DCTune calculation for intra-matrix to save some time.

After making the switch I've had some really good results with DCTune on 3 full movies (1 very clean but had some really complicated scenes in it, and 2 1940's movies that have alot of noise).

Did 4 encodes of each movie all had basic .avs except this was added for very light filtering of blocks/movement:
Undot().Temporalsoften(2,3,3,mode=2,scenechange=6)

Used MPEG, H.263, HVSBest, DCTuned mixed (MPEG intra, Q1 inter)matrices.

XviD settings were pretty much default except for:
1/1.25/1.00 B-VOP's
Trellis Quant
2-3 I-VOP quants
2-4 P-VOP quants
2-4 B-VOP quants

But like every matrix I've ever used it has advantages and disadvantages this is what I've noticed so far.

Advantages:
- High detail in the foreground (as much as HVSBest) while being more accurate and less smudges/white dots (not many people notice this but my eyes see it and its one of the most annoying things)
- Handles excessive film grain alot better than HVSBest (HVS-Best, H.263 blurs alot) and just as good as MPEG however I don't see a bunch of stray pixels like I do with MPEG.
- Least amount of block noise out of the bunch even better than H.263

Disadvantages:
- Background was the blurriest out of the bunch which can be a good or bad depending on the situation, I don't really mind it at all becuase it prevents alot of artifacts but I can see some people caring.
- Doesn't handle things that need very fine details well. That would include smoke, surface water, etc. However if it is in the extreme foreground it handles it just as good as MPEG. If its in the background it gets really blurry and macroblocks are apparent. This also doesnt bother me much considering I dont see alot of it in most movies and is a common MPEG problem.

Thats about all I can think of right now, overall I'll continue to use it with the MPEG intra-matrix. I think of it as kind of a combination with HVSBest detail in most situations and H.263 video stability. Which is what I've been looking for and haven't been able to find until now. But I was kind of disappointed using the intra-matrix DCTune creates but very pleased using the MPEG intra-matrix.

Thanks for the ideas and script and unlike most people replying in this thread think. I think it has alot of potential as being very good for creating custom quants for every different movie but it does need a few improvements, the MPEG intra-matrix is a pretty good improvement from what I've seen and alot of custom matrices use MPEG or something very similiar to MPEG for the intra-matrix.

SoonUDie

2nd May 2004, 12:26

@Colyptic
Interesting suggestion, I'll throw a matrix like that into my comparison. I'll also try some other mixes (HVS intra, HVS/DCTune blended inter, etc.)

Manao

3rd May 2004, 11:53

There is no point at all using the result of DCTunes in order to build an inter matrix. DCTunes works on pictures, while inter matrices are applied to the result of the motion compensation.

The statistical properties of a picture and of the result of the motion compensation are completely different. The fact that results are / seem better is only a matter of luck, or at best inadaptation of the HVS-Best matrix in that particular case.

Soulhunter

3rd May 2004, 19:36

Originally posted by echo-kk
This is what DCTune delivered after several hours :D of computing...
DCTune(1,1)50 Matrix

8 8 8 8 8 10 17 162
8 8 8 8 8 8 15 157
8 8 10 10 10 13 38 255
8 8 10 15 25 41 232 255
8 8 12 37 99 255 255 255
13 11 42 181 255 255 255 255
98 116 255 255 255 255 255 255
255 255 255 255 255 255 255 255

16 16 16 16 16 19 35 255
16 16 16 16 16 16 29 255
16 16 19 19 20 27 76 255
16 16 20 30 50 82 255 255
17 16 23 75 199 255 255 255
26 22 84 255 255 255 255 255
196 231 255 255 255 255 255 255
255 255 255 255 255 255 255 255

Several hours of computing for something like this... :eek:

Ok, I haven't tried thus tool neither read the doc's !!!

But what exactly is the goal in doing thus stuff... :confused:

Seems it simply remove all details by using 255 values !!!

Bye

SoonUDie

3rd May 2004, 23:17

@Soulhunter
Most of my matrices don't look like that. echo-kk must just have a source with a low amount of detail. The whole point of using something like DCTune is that it should be able to adapt to the source... high-detail sources will have less 255s, low detail sources will have more (supposedly removing what can't be seen anyways, so that you get higher quality elsewhere). One thing I've noticed about DCTune and DCTune-HVS matrices is that they have far less blocks than other codecs (though DCTune on its own is pretty blurry as well). Full comparison will be coming soon (my last finals are on Wednesday).

Manao

3rd May 2004, 23:33

Every matrices obtained by DCTunes posted here had as much 255 as iradic's one.

Colyptic

3rd May 2004, 23:49

@Manao:
Yes I can understand your point on it not being able to be used as a inter-matrix but have you read any documentation on DCTune?

I know its mainly talked about being used for JPEG compression. But if you look at the NASA press release of DCTune @ http://www.amestechnology.org/ames/tos.asp?NAME=dctunenew.html
they mention in a few places that it has benefits in MPEG compression, also if you click on the patent links on that page it goes in further details on its algorithm and usages.

The Intra-Matrix it gives is flawed I think when working with moving pictured though and its only detailed to a point but cuts out alot of very fine details which is why I think the results I was getting with it were kinda disappointing. However using a more detailed clip as a source (MPEG intra-matrix I chose) to predict motion is very important.

Using the DCTuned cutoff for inter-matrix has a very nice effect in my opinion. You aren't wasting a bunch of bits on things you normally wouldn't see and using the bits for things the human eye notices more, just like lumi-masking and contrast-masking. The HVS matrices do a good job of this however from my experiences they've given some pretty poor results in some movies. Main issues are movement in the background and sloppiness (stray pixels mostly) in the foreground. Taking a few minutes to DCTune each movie does a very good job of correcting this from what my eyes see.

After doing 6 full movie rips now each with the 4 matrices I mentioned earlier I'm still getting the best looking results from DCTuned matrices and they do vary quite a bit from movie to movie unlike any of the other matrices. So I don't think luck has anything to do with it.

As for mixing the HVS-Best matrix and just adapting to it I think isn't really much of an adaptation in most cases because if you look at the curve the DCTune varies alot from HVS-Best when using the first normalizing/averaging technique. The second averaging technique does produce matrices more identical to HVS-Best and I see inferior results using this type of matrix. I haven't tried this but I'm assuming it can be used effectively without mixing the HVS-Best matrix and using higher perceptual error values to get a reasonable compression level, doing this lowers the extreme right numbers and raises the extreme left numbers so it becomes more like HVS matrices as you raise the perceptual error while also lowering PSNR. Actually every encode I've done has had the same PSNR results in this order from highest to lowest: H.263, DCTuned w/MPEG intra, MPEG, HVS-Best. PSNR doesnt mean much to me (H.263 is inferior to HVS-Best in most cases in my opinion) but I guess some people are concerned about it since its mentioned alot.

Oh, can someone tell me what code needs to be changed in the script to not calculate the intra-matrix and use a constant one instead?

Thats all for now, bye.

Manao

4th May 2004, 01:00

Originally posted by Colyptic
have you read any documentation on DCTune?I read both links, and they don't say how it could serve in a MPEG scheme, except by using it only on intra frames.
Originally posted by Colyptic
The Intra-Matrix it gives is flawed I think when working with moving picturedHowever, it's the only matrix DCTune can properly compute.
Originally posted by Colyptic
However using a more detailed clip as a source (MPEG intra-matrix I chose) to predict motion is very important.I don't understand what you mean.
Originally posted by Colyptic
You aren't wasting a bunch of bits on things you normally wouldn't see and using the bits for things the human eye notices more, just like lumi-masking and contrast-masking.The point is, psychovisual heuristics used by DCTunes are meant to be used on real pictures, not on something that almost have the properties of noise like the result of motion compensation. In that case, there are no notion of contrast or lumi masking, because it's the difference between two pictures. The matrix you're computing simply isn't used inside a MPEG compression scheme the same way it would be inside a JPEG one, hence it is not pertinent. It may be better, because current matrices (hvs, mpeg... ) can be suboptimal ( no one knows ), but it's mere luck if it is so, and in that case there is no need to recompute each time the matrices. Just use the same one each time. It will have the same pertinency.

Originally posted by Colyptic
PSNR doesnt mean much to me (H.263 is inferior to HVS-Best in most cases in my opinion)Here, we're agreeing :)

PS : just to be sure you're not inverting concept : intra = key frames, inter = P / B frames.

Colyptic

4th May 2004, 07:14

@Manao

I'm done arguing about if it works on inter or intra matrices its not getting anywhere and I'm sure other people aren't appreciating it at all. I just saw your first post in the thread and say its useless and decided to say something, bad idea on my part. And please lets end it and get some constructive criticsm in here.

But heres my final words on inter and intra matrices with DCTune and if its useless like people are stating.

Don't know if you saw this but the press release states:

"The DCTune quantization matrix is compatible with industry compression standards for digital image compression such as; JPEG, MPEG and CCITT H.261." (Just mentions its compatible with MPEG but since they mentioned it I'd think it could and has been used and is in use by a few here)

and under benefits it states:

"Custom quantization matrices optimized for specific applications; printing, internet, web-TV, medical imaging, TV, digital video disks (DVD), digital video camcorders (DV), digital TV, high-definition TV (HDTV), teleconferencing,, etc."

Maybe they are saying that it only benefits MPEG in intra and you have to use something completely different for inter, but I doubt it!

Anyways, you say it won't work but have you tried it? If so, why don't you give some constructive criticsm that could possibly help improve it rather than say its not going to work. If you haven't tried it, then I gotta ask why are you saying its not going to work?

What I meant by using MPEG intra matrix with a slighter curve rather than a DCTuned one which is an extremely sharp curve is you get a broader range of information to use for the predicted frames, they are also alot more compressible than DCTuned intra matrices which alot of the time are alot of 12-14 in the upper-left area.

If you search those patent links for luma and contrast masking you'll find information on how it effectively does the same thing.

Using the same matrix for everything has given me not quite so pleasant results since switching over to XviD 1.0 builds about 2 months ago I've been kinda disappointed with HVS-Best (which I used nearly all the time with DevApi3 and happy with the results) on some of the encodes of really clean sources and tried H.263 (in which I never liked in DevApi3 but looks alot better in XviD 1.0 builds) and it looked preferrable over HVS-Best to me, and honestly never know what to expect when using one over the other. But my 7 different results with DCTuned have given me a steady result and will continue to use it until I see differently. I still believe that alot of movies can benefit more from specific matrices.

Yeah I'm aware of intra being still pictures (keyframes) and inter being predicted frames that rely on the previous keyframe. I may have made some typos or missed a few words here and there, it happens to me alot, I was never good in english class.

Thats all I have to say about will DCTune work with MPEG or not, my results say it does, yours may not but everyone thinks a little differently, Alot of people are bashing it without giving it a real chance.

SoonUDie

4th May 2004, 09:40

Of course, one of the problems with DCTune is that since we don't really know what's going on behind the scenes, it takes a lot of effort and experimentation to get good results. What may work well for one person might not for another. Only by continuing to experiment may we come up with a consistent positive result.

Manao, thank you for the information you have provided; your look at the theoretical side of things is valued. However, at the moment I think experimenting and comparing results still has something to add. In the end you may be proven right, but until then I think it's important to not give up.

M7S

4th May 2004, 12:47

Correct me if I'm wrong, but isn't there an easy (but time-consuming) way to check if DCTune intermatrices make sense?

Take four (or more) sources and make the inter matrices and combine them with the MPEG intramatrix. Use all four matrices on all four sources. You end up with four encodings per source. If the best encodings for all four sources are the ones, where the encoded source uses the intermatrix from that source, then its unlikely coincidence?

I'm sorry for that last sentence, my English could be better. But you do understand what I'm trying to say?

I hope someone has the time to make this test.

Colyptic

4th May 2004, 20:23

@M7S:

I might get around to doing that in that in the near future, it sounds like a good test. However from looking at the various DCTune (MPEG Intra) matrices I've seen from the 8 encodes (all 1CD) I've recently done, the results will probably either be very different or very similar depending on the sources used and how your eye sees things will also have an impact. Heres things I can think of that will affect the test:

- NOISE: 3 old/noisy sources tested and in the lower right of the table the numbers were alot lower (no 255's) compared to 5 clean sources. Also the curve was alot smoother (e.g jump from 40->50 instead of 40->255)

- DETAIL: Tested one blurry Pan & Scan DVD, it had about 75% of the matrix table at >100, which resulted in 640x480 DCTune vs. 576x432 HVS-Best/H.263 vs. 512x384 MPEG all right around avg. quant 3. This is the most obvious test I've done and DCTune definitely looked the best. Every other source has been fairly detailed and less than half of the table has been >100.

- HORIZONTAL/VERTICAL PLANES PRIORITY: DCTune matrices vary in amount of information it gives horizontal and vertical planes. Six of the eight encodes I've tested had priority on the vertical plane, one had a rather large horizontal priority and other was very close to being linear (Tested a marital arts movie that took place mainly in a straw field and DCTune very heavily prioritized the vertical plane, others have had about a difference of 10 in the step prior to 255 just to give you an idea). Other matrices don't take this into account. A few matrices prioritize vertical over horizontal and vice versa but how do you determine which one to use? You could get an educated guess by watching the movie beforehand but that may or may not always be right. Handling it linearly, like most matrices, more than likely results in bits wasted on horizontal when the vertical really needs it and vice versa.)

Thats all I've seen and if I do the tests I'll let you know the real results that my eyes see. I would include screenshots however I don't think its a very good comparison considering how it would normally be viewed in motion and dont have any webspace nor the connection to host clips here.

Asmodian

4th May 2004, 21:00

I just wanted to clear this up (for myself as much as anyone else).

The intra matrix is the matrix used to quant the DCTed picture information, this is all of the picture information - even the new blocks in p/b frames that couldn't be found in key frames and not just the picture information for key frames.

The inter matrix is the matrix used to quant the motion vector information. This matrix is never applied to texture type information such as the new macro blocks that couldn't be predicted with motion vectors from key frames.

The way the intra matrix is applied makes some degree of sense, the bottom is the very fine vertical detail and the right is very fine horizontal detail but I don't really understand how one would apply a matrix to motion vector information. The lower right is smaller vectors? What is the difference between the left and right edge? The top and bottom?

Sorry if this was obvious, I am just making sure I have the right idea as to what these matrices are doing (and how they are doing it is interesting).
Thanks for any extra info/corrections.

Manao

4th May 2004, 21:14

Asmodian : inter matrices aren't applied on motion vectors, but on the result of the motion compensation ( which is the difference between the current frame and the previous frame whose blocks have been moved by a vector ).

Motion vectors are not quantized.

Asmodian

4th May 2004, 21:49

Ah, I see - thank you very much (makes much more sense too).

So your point about the results of DCTune having nothing to do with that information is because it is difference information not actual picture information and as such should be very different from the actual picture information used by DCTune, except in rare situations?

Given this, the best way to use DCTune would be to use one of H263/MPEG/HVS inter matrices and an intra matrix determined by using this DCTune technique on all of the key frames from a first pass using a standard matrix? Maybe weighting key frames with more p/b frames more?

I will try some experiments along this line as soon as I have time (unless there is a better way to try?).

708145

21st May 2004, 16:11

Originally posted by Asmodian

So your point about the results of DCTune having nothing to do with that information is because it is difference information not actual picture information and as such should be very different from the actual picture information used by DCTune, except in rare situations?

Given this, the best way to use DCTune would be to use one of H263/MPEG/HVS inter matrices and an intra matrix determined by using this DCTune technique on all of the key frames from a first pass using a standard matrix? Maybe weighting key frames with more p/b frames more?

My guess is that using a DCTune derived intra matrix together with a modified inter matrix is worth testing.
Once I figure out how to install and _use_ this DCTune script I'll test by myself but I guess some of you guys are quicker.

For inter take the inter part of the matrix you would have otherwise used (HVSbest, MPEG, ...) and modify those fields where the DCTune derived intra matrix is >200 (200 being an arbitrary threshold).
The geometric mean springs to mind.

The idea is that if some frequencies occur very sparsely in the source (>200) then the occurance of those frequencies in block-diffs is mostly due to artifacts. In other words they occur less frequent than in hi quality movies.

If this works then derive a formula to adapt _every_ value in the original inter matrix depending on the respective intra matrix value.

708145

P.S. Hopefully it's a worthy first post. ;^J

yidaki

29th June 2004, 08:21

about the extract.bat
i dont have divx3 installed, and xvid wants to have some stats file, can i use any other codec instead?

DarkWave

29th June 2004, 09:56

...just go to your xvid-codec-configuration and set default settings.
after that start extract again...

it's the last stats-file, stored in ther registry, extract.bat tries to find. by setting default settings, the problem is gone

DW

ps.:sorry for my bad english

Sharktooth

29th June 2004, 12:19

DCT tune is a bit aggressive in distributing coefficents and generates one matrix only.
I'm working on a more complex temporal and spacial frequency analysis code that allows not only to "guess" the intra-frame matrix but also the inter-frame matrix.
All the process goes thru the analysis and motion estimation of key scenes or the entire movie (.avs input) and returns a custom matrix that should be (if there's no bugs - but there could be a lot...) HVS compliant.
From that point on, some parameters should be modified to obtain a high/low bitrate matrix.
Coefficient scaling is approximated around the HVS curve so coefficients may be rounded up or down.
However all this stuff is in a very very very alpha stage.

Asmodian

29th June 2004, 21:14

Originally posted by 708145
The idea is that if some frequencies occur very sparsely in the source (>200) then the occurance of those frequencies in block-diffs is mostly due to artifacts. In other words they occur less frequent than in hi quality movies.

Can you be sure of this? What is the difference between the predicted block from the previous frame and the block in the current frame? Can you be sure about anything to do with this difference? I don't have a good understanding of what frequencies these errors would be in or what the difference block would look like, but I don't think it would look much like either of the original blocks. Maybe it would be the block minus it's self moved 1/16 of a pixel (motion vector error?) with a few pixels also slightly different due to a very minor change in lighting during whatever motion is being predicted? In noisy sources the noise would also be different. Are only frequencies present in the original in the diff block? Would one want to loose the high frequencies in a difference block because they represent small differences or…?

Is there a way to visualize these blocks? Could we run DCTune on a set of difference blocks from P/B frames (I know this is getting a bit crazy)?

Sorry about all the questions, this is very interesting but I don’t really have time to do lots of testing myself and I would like to get a feeling for how this works.

@Sharktooth
Very intriguing! That would definitely speed up testing ;)