an idea for better motion and shape detection/compression ?

[toraneko]

I'm just wondering...
the bigger the size of a frame, the better/more precise motion and shape detection, right ?
could it then be profitable (not sure it is an English word) to add a function into the XviD GUI enabling to resize inside the XviD coder in place of AviSynth (I mean the "last" resize, at the last line) ?
that way, the coder would have more information to deal with, and it still could resize afterwards.

[ultimatebilly]

I don't know if this would work...
The problem is, that you divide the picture into blocks and macroblocks.
If you resize afterwards, you have new blocks and macroblocks, and the information in the picture changes.
That is the reason you cannot resize a movie without reencoding...
I don't know how resizing down exactly works, but I think it probably works like interpolating, with averaging between pixels and merging pixels...
If you imagine that, you see how much the information changes.
So a XviD-codec with your technique would gather the information about the unresized picture, and would then resize and compress (because compress and resize is not possible). I think the result is garbage, because the motion was detected for a quite different picture...
But I don't know nothing, it's just what I'm thinking...

[toraneko]

you "don't know nothing" ? so you know everything (or at least something) ! ("- * - = +")
just kidding :p
by the way, strange as it can be, that still IS the correct form in French ("je ne sais rien") ! who said French was an easy tongue ?

[ultimatebilly]

Not me! I always had great problems in school learning it, but now I wish I had invested a little more effort...
The grammars seemed always so unlogical to me, but that's probably what a native french speaker feels about german...
The most confusing thing in french for me are numbers.
Things like sixty-thirteen just sound weird to me...
But I guess all it would have required is a little bit of learning.

Now I'm not even able to ask for the way anymore... :(

By the way, my english-teachers always disliked the double-negative very much, but I think it sounds cool... (I know that it is wrong of course...).

[toraneko]

about numbers, there's still the way to speak like in Québec, Suisse (Switzerland) Romande or "old" French : "septente" = 70, "octente" = 80, "nonente" = 90 (sorry if the orthograph is wrong)

[trbarry]

For some time now I've been in favor of a possibly related idea involving resizing in the codec. The problem is that it would not be compatible with anything else in the world.

The idea would be to have a target quant that we don't go above. For any entire group of pictures that would be higher that this quant we would resize the frames down as needed. So different GOP's would be encoded at different resolutions. It would have to be done at the GOP level so it didn't mess up motion comp. They'd be scaled back up during decoding, where the target resolution could just be the display resolution. Practically everything get's scaled at least once these days anyway.

We've known for awhile now that to get certain bit rates we can't exceed certain picture dimenensions and it is better to resize the whole clip downwards if we'd otherwise end up with all macro blocks. So this would be using resize as a means of rate control. Fast action detailed scenes would then end up encoded a lower resolution, determined by seeing a high quant. All this could probably even take advantage of hardware scaling for PC playback.

But I guess there is no way to really do this and still be valid MPEG-4.

- Tom