another tempgauss postprocessing idea

[Terka]

After clp.TempGaussMC_beta2()
there could appear big areas which move by cca 1 pixel
eg. frames 11->12 on psi clip.
Should be: whole area should move at once.
Is: area is separated into parts that move in different times
Idea: use corrector and shifted frames, to patch the wrong timing
Result of following code: not good.
Is the idea total crap?


#tg=tempgaus
a=tg.FFT3DFilter(sigma=10,bt=5) #this is calm, but blurry

shift=[<"shift", 0, 2, 0.1>]

a1=tg.Spline16Resize(tg.width(), tg.height(),-shift,0)
a2=tg.Spline16Resize(tg.width(), tg.height(),shift,0)
a3=tg.Spline16Resize(tg.width(), tg.height(),0,-shift)
a4=tg.Spline16Resize(tg.width(), tg.height(),0,shift)

blur=[<"blur", 0, 1.5, 1.4>]
a1b=a1.blur(blur)
a2b=a2.blur(blur)
a3b=a3.blur(blur)
a4b=a4.blur(blur)

tgb=tg.blur(blur)


CorrectTh=50
out=corrector(a,a,tgb,tg, a1b,a1, a2b,a2, a3b,a3, a4b,a4, Mode = 1, Th = CorrectTh)

[Terka]

FFT3DFilter(sigma=10,bt=5) calms the nervous areas described in previous post.
the nervous movements are +-1pixel MCompensate should move the pixels to correct position.

#tg is result of tempgauss
tg=avisource("c:\TempGaussMC_beta1.avi").ConvertToYV12().Crop(left,top,-right,-bottom)
tgC=tg.FFT3DFilter(sigma=10,bt=5)

thSCD1=5000
bs=32

dva=Interleave(tg,tgC)

source_super = dva.MSuper(pel=2, hpad=16, vpad=16, levels=1) # create the source superclip
calm_super = dva.MSuper(pel=2, hpad=16, vpad=16) # create the calm superclip

bv1 = MAnalyse(calm_super, blksize=bs, isb = true, delta = 1, overlap=bs/2)
fv1 = MAnalyse(calm_super, blksize=bs, isb = false, delta = 1, overlap=bs/2)

MFlowInter1=dva.MFlowInter(source_super,bv1,fv1,time=50.0).subtitle("MFlowInter1")


f1=dva.MCompensate(source_super, fv1, thSCD1=thSCD1)
b1=dva.MCompensate(source_super, bv1, thSCD1=thSCD1)

f1=dva.MFlow(source_super, fv1, thSCD1=thSCD1)
b1=dva.MFlow(source_super, bv1, thSCD1=thSCD1)


out=f1.SelectEvery(2,1)

#out=MFlowInter1.SelectEvery(3,1)
stackhorizontal(tg,out,Overlay(tg,out, mode="difference", pc_range=true).tweak(sat=3.75))
Didee, could this work?

[Terka]

with this method i was not able to get good resuts :confused:

[Terka]

found
ythresh=25
cthresh=50
maxdiff=206
LUTDeCrawl(ythresh=ythresh,cthresh=cthresh,maxdiff=maxdiff)
effective on dancing grass. without much blurring.

[markanini]

Why not upload some clips, Terka? Reducing temporal artefacts while retaining details is always interesting.:p