Consecutive INTRA frames

[octapus]

I am encoding at the moment Rush Hour 2. It's about 90 minutes movie and i want it in one (1) CD with ogg vorbis sound.

Well i encoded the movie with modulated quantization type on 2nd pass and i noticed something strange( i look at debugview to see how encoding performs)

1) At some points of the movie there are consecutive INTRAFRAMES. And i didn't find just one. I found a lot of them. the maximum was 8 consecutive INTRAFRAMES at about 28.000 bits. I remember i found a topic talking about this but i can't find it. It was someone who stated how he fixed it.

2)The above encoding was done with the external way (GuardianKnot). I put a weight of 2% in credits and therefore the rest of movie gets a little more bits. By this way when i encode with modulated quant. type i get a lot of frames with quantizer <=3 and therefore a lot of MPEG type frames. Of course everything above 3 is encoded with H.263
P.S Credits are completely readable.

But when i try to encode with internal way(where again i encode credits with 5%) there are very few frames <=3 and therefore almost all frames are h.263 qunantization type?

It seems that the internal routine is compeletely different.
Visually GuardianKnot's way gives better results according to my opinion. What do you think?

[rui]

Originally posted by octapus

It seems that the internal routine is compeletely different.
Visually GuardianKnot's way gives better results according to my opinion. What do you think?

Really? I must try this.

Back to testing stage ;)

But using Gnot to calculate the curve doens't allow for the automation that Vdub allows with the internal 2 pass. :(

[-h]

I'll add some code to restrict subsequent INTRA frames in the next commit. I'm not sure if it's a good idea, mainly because those INTRA frames are usually smaller than the INTER frames that we would be replacing them with. That is just how the motion estimation works at the moment.

If GKnot is assigning some frames a quantizer <= 3, it's taking those bits from somewhere - most likely action scenes. Depends on what kind of frames you prefer details in. Hopefully, I'll be adding Foxer's Nandub-esque curve compression / payback code to the 2-pass modes this weekend, so both internal and external 2-pass encodes should get "better" overall.

-h