Comparing Codecs for 2004

MunchMunch writes "Popular encoding/guide/news site doom9.org has just put up its codec shoot-out for 2004, comparing 3ivx 5.0, Divx Fusion 5.9 (prerelease 6.0), Nero Digital Main Profile and High Profile, RealVideo 10, On2 VP6, VideoSoft's VSS, Xvid 1.0, MS's WMV9 and, last, newcomer Jomingo's HDX4. The comparison covers the speed, accuracy, target-file-size-adherence and other aspects of the codecs -- but also lets you compare yourself via high- and low-bandwidth framegrabs of each codec with a nice zoomable image-swap script."

Comments

Theora?

[mano78]

From a quite-newbie point of view: is there a reason why Ogg Theora isn't included? Given the quality and increasing popularity of Vorbis, I would have expected at least a mention. And it would have been interesting to know its state relative to the others.

Re:nobody seems to have heard of this codec

[Elledan]

Valid link to the thread about the Snow wavelet-based codec: Snow

For those too lazy to click the above link, here's the content of the first post:

"I think a new thread is a more fitting place to discuss about the Snow codec. :) If someone wouldn't know what is it, it's an experimental wavelet codec made by the ffmpeg developers, which borrows a lot of tools from h.264, and while it's still early in the development, it's already giving very good results, far surpasses other wavelet codecs (rududu, dirac) and imho Xvid too, quality-wise. Unfortunately it's only usable with mplayer/mencoder right now, but i think the next ffdshow will include it, so the testing will be more easier. [Update: The latest ffdshow build provided by Celtic_druid have Snow support]

I've played with the settings, and so far this command-line gives the best result:

code:mencoder in.avi -o out.avi -ovc lavc -lavcopts vcodec=snow:vstrict=-1:vqscale=3:qpel:v4mv:cmp=1:s ubcmp=1:mbcmp=1:pred=1

This gives ~600-800 kbps, depending on the source, and the quality is excellent imo.

vqscale is the quantizer, if it's not included it in the command line, Snow will compress losslessly.

So far my opinion about the different settings: qpel always increases the quality - recommended v4mv - i would only recommended it at lower quantizers (max 4-5), above that the stronger artifacts it causes like ringing can hurt the quality xxxcmp=1 (using SSE comparison method instead of SAD) slows down the encoding, but prevents the color mismatches, which can occur otherwise (anyone who tried rududu codec can remember to that). using pred=1 or 2 (different wavelet functions instead of the default) can increase the quality, but these make the encoding (and pred=2 the decoding too) much slower."

Wavelet-based encoding definitely sounds like a great idea. It's only too bad that it isn't universally usable (it can't compress certain images well, either), and requires a fast CPU. At least it gives that Athlon 64 3500+ you just got something to do :)