FFVFW testing results/concerns
 

WELCOME TO FFVFW WORLD!!!!

I have decided to post a new thread that only relates to FFVFW, as the other thread seems to be foucsed upon MEncoder exclusively.

I believe that I have solved a problem that I ran into with blocks created around low bitrate areas within frames. Notice that I did not say low bitrate scenes, as ffvfw sections off areas within frames and applies a heavy compression. At low bitrates this is not a problem, but for the higher bit rates of DVDs, it proved to be too ugly for me to stand. The setting of spatial complexity masking seems to solve this, higher settings create less blocks. Confirmation of these results will be posted sortly tonight...

Also, one of the advantages of doing a two pass encoding is being able to tailor the final file sizes to an appropriate lenght that matches the amount of data used up by the stream. For example, for BluePlanet episode 1 first pass ended up being around 1,300,000 KB in length, while BluePlanet episode 2 ended up being only 502,000 KB in length. Rather than having both of the files encoded at, say, 800 MB, would can tailor the file sizes to give extra data to the much large episode 1. That way constant quality is preserved, rather than adjusting the Q level in CCE or TMPGence (for example).

Hi Poeschr,

I'v experienced the same 'ugly' blocks you ran into (high and low bitrate encodings). Both Mencoder and ffvfw gave me those blocks on flat survaces. A post by Marcellus gave me a direction in which to search for a sollution. Marcellus wrote that he thought that the blocks were due to 'over quantizing' of flat survace's and that he wished he could use min_quantizer=1. With mencoder I tried it and no blocks! But maybe more interesting is the fact that when I turned of 'trellis' the blocks were also gone. I believe trellis is doing a to good job in quantizing (over quantizing) the frames. The filesizes went up whithout trellis ofcourse....

Maybe you could try this with your settings in ffvfw (I only tried Mencoder sofar. ffvfw is so sloooooow compared to Mencoder.)

Tinus

You bet staigerpaip :!:
I wonder why since both are based on the same code :roll:
Hope we can find that out with Milan's help.
Cheers

Mencoder is compiled under the GCC ports MinGW/Cygwin, while FFVFW is compiled under Visual C++. And I read a while ago the FFVFW developer (forgot his name) saying that libavcodec (the commom part of mencoder and FFVFW) gets a lot slower when compiled in Visual C++.

I have to specify the 'ugly' blocks a little bit more. My findings are the combination of trellis with low bitrates. With dvd specs there are no notible blocks. With svcd specs (min-300, max-2500, avg-1000) combined with trellis the blocks are VERY notible on flat surfaces.

Try this on FFVFW as a replacement for trellis, it's faster and may work better:

Input -> Enable Image processing -> Blur & NR -> denoise3d 4,3,6 -> HQ

And this solved me a B-frame problem is a nasty anime stream I have (nasty = interlaced, blended fields, high motion, very complex scenes)

Motion Estimation -> Me pre-pass -> Allways

Note: my feedback comes from mencoder testings.

Bilu

Yes, I know. I follow your posts in the mencoder tread close. Very interesting. I'm not testing ffvfw anymore, just mencoder. I also abandoned trellis. To slow and blocky (I'm testing SKVCD). Thaks for the tip on the prepass. I also have difficuties with the b-frames (pictures is not solid between b-frames. I get see-sick watching it :wink: ) If I have some results, i will post them in the mencoder thread.

Greet's,

Tinus

yes testing confirms it..turning off trellis q reduses the blocks...

Did some testing with "rate distoration.." found large blocky scenes at highbitrates...recommend using the default value of "fewest bits" - especially at DVD resolutions