"Curve compression" just means taking bits from high bitrate scenes into low motion ones. It's more or less like what you describe, but on whole scenes instead of individual blocks. It's called like that because if you look at a bitrate per time graph, it will get compressed (taken to the extreme, it would become a flat line: at that point you are doing constant bitrate). Its advantage is that it's very simple and therefore it's implemented on a lot codecs, but wouldn't help in the example you mention.
I'm glad you're going for the per-block route, it can be quite the gain. I assume there's no estimation on when it'll be ready, right?
If I read your post right I think you got some details about x264 wrong (not that I'm an expert either). x264 has a lookahead (--rc-lookahead) much bigger than the B-frame detector (--bframes). For example by default you have at maximum 3 consecutive B-frames, but the lookahead is 40 frames. To decide frame type it just considers 3 frames at once and uses either heuristics or an exhaustive search at high options (--b-adapt 2) which is very slow for >6 or so frames. Separate from that, the lookahead (a low quality, reduced resolution encode) aids both in ratecontrol (specially useful with buffer constraints) and in computing block value (what you are doing). It certainly doesn't need B-frames to work, it works fine in Baseline profile which doesn't support B-frames (in that case the B-frame detection code doesn't run). Additionally, on two-pass mode, it writes the block values to disk, getting effectively infinite lookahead. Were not for the (large) lookahead, it would only work effectively in two-pass mode.
Re: Ghost, Ptalar and demos
Date: 2010-03-15 10:37 pm (UTC)I'm glad you're going for the per-block route, it can be quite the gain. I assume there's no estimation on when it'll be ready, right?
If I read your post right I think you got some details about x264 wrong (not that I'm an expert either). x264 has a lookahead (--rc-lookahead) much bigger than the B-frame detector (--bframes). For example by default you have at maximum 3 consecutive B-frames, but the lookahead is 40 frames. To decide frame type it just considers 3 frames at once and uses either heuristics or an exhaustive search at high options (--b-adapt 2) which is very slow for >6 or so frames. Separate from that, the lookahead (a low quality, reduced resolution encode) aids both in ratecontrol (specially useful with buffer constraints) and in computing block value (what you are doing). It certainly doesn't need B-frames to work, it works fine in Baseline profile which doesn't support B-frames (in that case the B-frame detection code doesn't run). Additionally, on two-pass mode, it writes the block values to disk, getting effectively infinite lookahead. Were not for the (large) lookahead, it would only work effectively in two-pass mode.