How to make "short" GOPs with mpeg2enc?

[FredThompson]

Saturated red yields some horrible mosquito noise when it's used to make an I-frame.

mpeg2enc.exe -f 8 -o menu.m2v menu.yuv

The noise is lessened, dramatically, if B and P frames are present.

In this case, the I-frame is used to make a DVD menu background.

What if a short GOP is used instead?

A quick test of GOPs made by TMPGEnc, QuEnc and HC shows far less noise but they all make rather large M2V files.

Is there a way to have mpeg2enc make a very short GOP with all 3 types of frames?

The attached image helps to illustrate the differences. The source bitmap has anti-aliased text on a solid red background.

[mpucoder]

Would it be possible to get the source bitmap so I can test Muxman's internal I-frame encoder? Or maybe you could try it and tell me how it stacks up.

[FredThompson]

Sure. Here 'tis: http://home.mindspring.com/~fredthompson/custommenu.bmp

"Best" solution would be a very nice I-frame, quite frankly.

[mpucoder]

It certainly didn't look good with the default quant matrix, but I figured for a menu with text why not try minimum quantization. The result is at http://www.mpucoder.com/Muxman/custommenu.m2v

This was a custom compile, but I think I'll add this quant as an option.

[FredThompson]

Is that a complete file? It crashes VirtualDub-MPEG, MPC and PowerDVD.

Guess I'll be adding Muxman to my toolbox.

[mpucoder]

Something went wrong - it looked OK in preview, but the file is not good.

[FredThompson]

Well, you've got me eager, that's for sure. This whole saturated red/blue in 4:2:0 has been a royal pain.

[mpucoder]

I think it's fixed now - one computer is having a problem.

[mpucoder]

Forget that file - the idea is good, but I need to work on it a little. The problem is the file is too big for a DVD buffer. The maximum is 232K, and this frame is 270K.

[FredThompson]

What about lopping 8 pixels off each side? 704 width should be valid, right? That would help a little. Can't hurt.

Looks like I forgot to cetner the source image and the embedded text properly. Oops.

[mpucoder]

I changed the quant just a little and got it down to 186K. My software player just went nuts and won't play anything, not even known good, but the data looks OK in VobEdit.

[FredThompson]

186 is pretty good! As soon as I put the last reply up I realized it was a silly suggestion. The goal wouldn't be to come as close as possible to the limit, it should be good results and not go over the limit, even if the source is more complex.