Jun 28 14:23:32 T0B1A5, there are several ways of doing hierachical ME Jun 28 14:24:34 and they do work. however hierarchical ME is interressing only because it has an indirect effect on MV coherence Jun 28 14:24:51 imho it's much better to enforce that coherence by other means Jun 28 14:27:11 the two main alternatives for hierachical ME are : Jun 28 14:27:20 - using coarse to fine block size. (no speed gain) Jun 28 14:27:46 - using coarse to fine resolution. (ME searches on reduced resolution data aren't very usefull) Jun 28 14:28:22 ME is very simple as soon as you understand that the problem is ill-posed Jun 28 14:32:47 ill-posed? Jun 28 14:59:39 yes Jun 28 14:59:47 it's an undercomplete problem Jun 28 14:59:52 more unknows that constraints Jun 28 15:01:01 so to improve results, we have to add constraints based on our expert domain knowledge Jun 28 15:02:39 btw, i'm talking about natural (true) motion Jun 28 15:03:10 block based ME in video codecs is used as a vector prediction Jun 28 15:04:09 that's why stuff like RDO ME can make sense, as long as you have enough texture bits Jun 28 15:04:32 if the texture bits are too low then i believe RDO ME is harmfull Jun 28 15:06:54 (because keeping a high MV consistency will yield better visual results even if it has very high distortion) Jun 28 15:14:33 so realvideo actually does the right thing(tm)? Jun 28 15:14:39 for sub 100k video Jun 28 16:05:12 prunedtree: thanks for the answer about hierarchical ME Jun 28 17:17:38 so realvideo actually does the right thing(tm)? << why are you saying that ? Jun 28 17:17:57 they said they avoid rdo me for low bitrates ? Jun 28 17:18:38 btw average bitrate is meaningless. you could find very low texture data on many frames even at 2+ mbps Jun 28 17:19:54 well, for low bitrates the image distorts a lot but mv field is homogeneous Jun 28 17:44:56 maybe that's why bluesky sucks a bit in x264 Jun 28 17:45:02 i never checked the mvs