GOPs in TMPGEnc [CBR and VBR]

[Makaris]

I've noticed something odd about the way TMPGEnc handles GOPs and I wonder whether anyone can help me understand.

I encoded a short sample at 4900 kbs using CBR (constant bit rate) and found that the sizes of the resulting frames throughout the clip were typically :

I - 120 Kb
P - 30 Kb
I - 9 Kb

[This clip contained little variation in subject matter and fairly little motion. I left the GOP structure at the default, which gave me IBBPBBPBBPBBP (13 frames).]

I then encoded the same sample at the same rate [4900 kbs] using VBR (variable, 2-pass); the bitrate ended up pretty consistent at 4900 kbs throughout (as expected, since there wasn't much variation in the content), but I now found that the resulting frames were typically :

I - 65 Kb
P - 30 Kb
B - 15 Kb

Note that in CBR the ration of I-size to B-size is about 13:1 whereas in VBR it's less than 5:1. It looks as though VBR is giving much more weight to the motion content (by expanding the B-frames) at the expense of detail or image quality (by compressing the I-frames much more heavily).

This seems to me a very significant difference and I don't see why it occurs. As I understand it, the I-frames are critical in defining the amount of detail in the image, since B and P frames are based on that set of data. Moreover, my experience of using JPGs tells me that while you can compress a 720x576 image to 120 Kb without noticing any artefacts, going on down to 65 Kb degrades the image considerably..

Obviously, VBR gives compensating, benefits over a longer clip by allocating higher bitrates to certain sequences; but, on the face of it, this treatment of an "average" clip suggests to me that CBR is likely to produce better results than VBR here.

Am I missing something? By the way, I tried to tweak the algorithm by changing the "spoiling" values for B and P-frames - but that seemed to have no effect at all on anything...

[disturbed1]

In using a 2 pass VBR, TMPG does a first pass to give an average bitrate, and takes note of motion. During the 2nd pass it takes that information and attempts to maintain a certain quality (Constant Quality) acording to motion movement, and desired bitrate acording to quantization.

Since the clip had very little high motion sequences, TMPG decided to dedicate more bits to the B frames, which aids in fluid motion during still sequences, and slight panning.

Given a different sequence of video, the results can be quite different. If you look at a whole movie (2Fast 2Furious) which include both high motion and no motion, you'll see that TMPG does allocate bits more appropriately according to motion.

If TMPG's bit allocation doesn't suit your needs, you can tell it what to do to each frame. Including asigning I,P,B per frame, quality per frame, quantize matrix per frame, and bitrate per frame.

[Makaris]

Thanks for the reply - and for understanding what I meant despite my typo in a critical place [ I wrote "I - 9Kb" where I meant to write "B - 9Kb".. - which rather destroyed my point !! ]

I hadn't considered that B frames might be of more significance in "still sequences"