visual comparison of custom matrices on medium/low bitrates

[bond]

lo guys,

i invested some time and tested a whole bunch of custom quants matrices, using the matrix clip doom9 uses too on 650kbps (i think its a question of definition whether this should be called a medium or low bitrate)

i tested the following matrices against h.263:
very low bitrate, hvs_good, hvs_better, hvs_best, mpeg

to save bandwidth i packed the whole comparison + screenshots into a .zip you can grap here (http://maddxshow.sourceforge.net/download/matrices.zip) :)


over the thumb as a result of the comparison i would rank the matrices subjectively according to their output quality and of course heavily depending on the used settings/sample in this comparison this way:
hvs_best > mpeg > hvs_better > hvs_good > very low bitrate > h.263

looking at these results i think that, when speaking about 1CD encodes, this comparison shows that the so called "high bitrate" matrices are not only good for high bitrates, but also seem to be suitable for good compressable sources on low bitrates
the so called "low bitrate" matrices are for me more suitable for really hard to compress movies, like lotr, pearl harbour aso i think


coming to the end i want to say that custom matrics are really a very powerful tool in video compression and can influence the output very strongly! its really important to explore all the possibilites they offer
if i find the time i will test some more high bitrate matrices if they could be useful for a matrix 1cd backup :)

[Teegedeck]

Thanks for conducting this test! I'll have a look at it tomorrow. For now I must say that your results match my personal impressions, except that I gave hvs-better and -good a pass after rather short testing.

[SeeMoreDigital]

Yes, nice one bond.

I too will have a look tomorrow. I tried Soulhunters MPEG custom matrices with XviD to generate some HiDef material. I also used them with Recode2. The results were quite promising.

Sadly the test were only short, barely long enough to benefit from using 2 passes.

Cheers

[JasonFly]

I have done such a test some month ago and I hd quite the same results. I perfomed a test with two differents clips with and without noise and tested with différent matrix.I just remenbered that visually, the winner was HVSbest and the "looser" was H263. I don't remembered the orfer between HVSbetter/good and MPEG but that's not very important.

Since this test,I always use HVSbest, even with low bitrate rip and the results are good for me.

Thanks for confirming my point of view.

[SoonUDie]

I downloaded your samples, and I have to say I come to different conclusions. I don't have time to do a full comparison at the moment, so I decided to simply compare h.263 and HVS_Best. My method of comparison was layering the images in paint shop pro and toggling the opacities.

In general, I found that h.263 had the most pleasing images in many cases. The only time where I found HVS_Best noticably better was on scene 5 (closeup of agent smith with the other two in the background), where HVS simply dominated h.263 in terms of detail. Scene 7 (shooting) was a tie in my opinion, though HVS_Best seems to have the advantage in terms of detail. I would have to see the full scene to decide. In all other scenes, h.263 and HVS_Best were pretty much tied, but in an apples to oranges way: h.263 preserves significant detail very well, and HVS_Best sacrifices some of that quality to add more detail in lower contrast areas. I ran the same comparison tests a while ago on different source material (Men In Black II) and the results were the same (though as a result, I came up with a matrix that's roughly half-way between the two).

Overall, they're not really that different in quality. Personally, I would love it if we could switch back and forth between the two for different scenes; h.263 for motion, HVS_Best for static.

From what I've seen here, and from my own tests, I would suggest HVS_Best to people who want the most detail possible in static scenes and don't mind a few artifacts in motion scenes or a slight loss of precision. However, if artifact-free encoding is what you're after, I think h.263 does a better job.

EDIT: I don't want to re-write this, but my feelings have changed. I've started testing matrices on sources I haven't used before, and h.263 is getting whooped. On the plus side, though, I may have developed a custom matrix even better than HVS_Best. Details will follow... eventually.

[Colyptic]

I've spent hours upon hours doing matrix tests using constant quant 3 (aim for this as an average every time) on 4 torture clips (snow scene, floating walls scene, dark/smoky scene, dark/smoky/floating walls scene) and 2 very clean clips using every matrix found in the pack LigH posted (Thanks!) and H.263. Used XviD 1.0 RC4 and XviD 24062003 (DevApi3) build. For resizers tested Lanczos (I use for everything nowadays but the alias problem can get a little annoying) and Neutral Bicubic (Good but makes complicated scenes like these look very blurry). But I'll just base this on the 2 matrices that are mainly being talked about (HVS_Best and H.263).

With DevApi3 build, HVS_Best was really high up on the list (Fox Home Entertainment at quant 6 to match size topped the list) and has been my choice of matrix for months now and by far the best compared to the matrices you tested but looked even better with a light TemporalSoften to calm down the movement. Whereas H.263 was inferior to most of the matrices tested. It had the common 'swimming' problem, looked very blurry and at times the whole screen was unrecognizable.

However, I switched to XviD 1.0 RC4 a week ago. After being kinda disappointed with the typical HVS_Best I used with DevApi3, I ran the same kind of tests yesterday with just a few of the more common matrices. To my surprise H.263 came in second only to Fox Home Entertainment. This H.263 compared to DevApi3 H.263 showed alot more detail, alot more precision with alot fewer artifacts (mainly moving background and there was no more noticable 'swimming' artifacts). HVS_Best showed no more detail to my eyes and had alot of movement in the background as well as was very sloppy (blocky, bleeding colors, ringing, and trails of pixelation).

Comparing the two that came out looking better (HVS_Best with DevApi3, H.263 with XviD 1.0) I chose XviD 1.0 RC4 using H.263 and Trellis quantization with max quant 4 for most 1CD rips. But if I'm able to get max quant 6 with Fox Home Entertainment I choose it.

Also gave blind tests to a few experienced rippers and people who know nothing at all about encoding to give me their opinion and the majority agreed with my results. I'm kind of wondering if something has happened in the XviD 1.0 builds that could cause these drastic change in results.

[Soulhunter]

Originally posted by SeeMoreDigital
I tried Soulhunters MPEG custom matrices with XviD to generate some HiDef material. I also used them with Recode2. The results were quite promising.Btw, the other matrix I posted there (the anime one) seems also to work well for real movies @ lower bitrates... ;)


Bye