Avisynth: FaeryDust filter is a time bandit!
 

I mean this in the nicest way of course. I did some tests on the encoding time increments caused by the filters in the Optimal script.

I used the script on a reasonably noisy capture from a home movie (340x480) and fed to Tmpgenc using KVCD 352x576. I started with:

LegalClip()
#FieldDeinterlace()
BicubicResize(352, 576)
#STMedianFilter(10, 30, 0, 0, 10, 30)
#FaeryDust()
#unfilter(50,50)
#mergechroma(blur(1.50))
#mergeluma(blur(0.2))
#Convolution3D (0, 32, 128, 16, 64, 10, 0) # VHS capture Bad Quality
LegalClip()

Sequentially I removed the comment lines from top down and recorded the increase in encoding time.

Results are normalised to a 1sec clip.

As above (no filters).. 3.8s
FieldDeinterlace......... 0.0s
STMedianFilter........... 0.5s
FaeryDust.................. 7.5s
unfilter....................... 0.0s
mergechroma............ 0.3s
mergeluma.................0.5s
Convolution3D.......... 1.0s

Fairy dust processing is twice that of basic resizing and encoding!!

I did not take any note of file sizes. I note also that the filter times can depend on the work done by the previous one, although, taking FaeryDust out did not affect the times of the filters following.

ozjeff99

Great data ozjeff99 :)
I would like to see that on a longer clip! Maybe 15 seconds will be more accurate, because one second is just not enough frames for a good average :idea:

-kwag

In conclusion, fearydust kills your encode so dont use it unless you have to. :lol:

Thanks Kwag. It was actually a 4 sec clip and I normalised result to more easily show relativity. Whilst doing it I did a quick check and it seemed to show linearity. But I agree it would be better to check on longer clip particularly to show more significant digits and give figures for those that are rounded to zero.

It's 1pm and its p*ing down in Sydney so I'll leave till tomorrow.

Cheers ozjeff99

Hey ozjeff99,

So was your testing done on a Pentium-1 166MHz PC with 32MB SD-RAM, or a Dual-P4 3GHz machine with 5 Gig of PC3200 DDR-RAM? :wink: Your test results will have even more meaning to people if they know your PC's stats. 8)

-d&c

Hi Dazed & Confused. Why do you speak for everyone else? I am showing the relative times (normalised) here not absolutes which, within reason, should be largely independent of hardware. I thought the rest of the universe may see that :wink: However in answer to your 2 questions the answers are no and err.. no. It is a standard P4, 1.6Ghz, 256RAM, 80GB cleaned HDD.

I might get the 5 Gig of PC3200 DDR-RAM this afternoon if you think its a good thing :)

Cheers ozjeff99

Kwag....here are the results with a 15sec clip (extension of original).

As above (no filters).. 55s
FieldDeinterlace......... 2s
STMedianFilter........... 6s
FaeryDust.................. 110s
unfilter....................... 2s
mergechroma............ 2s
mergeluma.................7s
Convolution3D.......... 16s

If you normalise these ie assume 15sec = 1u (time unit)

Clip lentgh = 1u

As above (no filters).. 3.67u
FieldDeinterlace......... 0.13u
STMedianFilter........... 0.40u
FaeryDust.................. 7.33u
unfilter....................... 0.13u
mergechroma............ 0.13u
mergeluma.................0.47u
Convolution3D.......... 1.07u

Compared to my earlier post, as can be seen results are reasonably linear with clip length although mergechroma fairs a bit better. I think there would have been some rounding errors since I was getting my times from Tmpgenc which are rounded to whole seconds without fractions.

Is there a better option than FaeryDust anyone?

Cheers ozjeff99

Maybe SpaceDust :idea:

Hi Kwag ..... for 15sec clip SpaceDust() adds only 5sec. to processing time.

Or normalised:

Clip = 1u (time unit)

SpaceDust......... 0.33u

ozjeff99

SpaceDust uses Spatial Filtering only, while FaeryDust uses Temporal Cleaning only, so I don't think you can substitute one for the other...

Yes, but SpaceDust is the fastest of all the "Dust" family, and it does a pretty good filtering job :)

-kwag

I did some tests on my clip which is of medium to low VHS quality and using AVSCompare (wonderful little prog) ran 2 clips side by side from same source file with 2 scripts the second with the dust filters. I found only marginal improvement with any of the dust filters. The one that did a beautiful job was Convolution3d which is both a spacial and temporal filter. I'm wondering if any of the dust filters really have a place with captures?

Hi Dazed & Confused

Thanks again for the information that you PM'd me.

Cheers forum friend 8)
ozjeff99