Revision as of 14:33, 29 October 2013 by Jack
# Meeting 2013-10-22 ## Agenda - even/odd quantizer behavior - determinent 1 search results (gmaxwell) - IETF - CFL using luma sse ## Attending derf, jmspeex, unlord, gmaxwell, xiphmont # even/odd quant zigzags >>- g: Anyone have any theories on the difference of the even/odd behavior with >>scalar quantizers. >>- no >>- jm: we may not be doing the rounding right >>- m: I can look at that. >>implying - m: Still intend to look at this next I think ## determinent 1 Greg dropped off mumble, can't seem to get back, ah there he is has filters for all the sizes, not run videos through them not trained intra stuff for 16x32; the training stuff is not working nathan has patch that does it in the codec, waiting for greg to review CG for raw transform, no prediction: 'I don't know yet' Does it quilt/block/do 'funny' things?: Checked 4x8, 8x16 'nothing too funny' but not quite a smooth-- not quilt/block just not quite as smooth. ## IETF IS it good to continue having Xiph/Moz there in force when we're not there to stack votes? :-D We don't vote in the IETF! In-duh-viduals And we don't stack either (yeah, all four of us) ...this never happened? TO THE TIME SLIDER - m: OK, established I'm going, need to get onboarding more in hand to book travel # CFL using luma sse http://people.xiph.org/~unlord/crfl_v_luma_sse_4x4.png "Chroma red from luma vs Luma" http://people.xiph.org/~unlord/cbfl_v_luma_sse_4x4.png "Chroma blue from luma vs Luma" X axis is prediction error (residual) in the chroma planes Y axis is prediction error (residual) in the luma plane All 4x4, no basis scaling scatter plots of summed squared errors for Chroma from Luma prediction vs simple luma prediction (fruits only) dude, mismatched axes dude, axes not labelled No obvious correlation between errors JM suggests looking at the goodness of the least squares fit as a hurestic in the bitstream for how good the CFL is working GM points out that at least in some cases there appears to be a dynamic range issue in the coeffs that aren't used in the training, so looking at the goodness of the training fit may not help there Goal is to come up with a CfL that, when it screws up, is not worse than nothing, or is seldom worse than nothing Lots of discussion of looking at optimal results vs the corner cases we're seeing.... Nathan is on it JM at AES saw a FHG booth directly going after Opus: http://jmvalin.ca/misc_stuff/AES_show_FhG.jpg They had a couple samples all at very low rate comparing Opus to USAC and prerelease versions of the 3GPP EVS.