DaalaMeeting20131112
Jump to navigation
Jump to search
# Meeting 2013-11-12 ## Agenda * CFL fixes ## Attending gmaxwell, jmspeex, nathan Tim, Xiphmont not in today (From Nathan's status report:) Fixed bug in od_resample_luma_coeffs, -6% on rate +0.3dB in PSNR: https://pastebin.mozilla.org/3586591 http://people.xiph.org/~unlord/head_vs_cfl-psnr.png http://people.xiph.org/~unlord/head_vs_cfl-psnrhvs.png http://people.xiph.org/~unlord/head_vs_cfl-ssim.png http://people.xiph.org/~unlord/head_vs_cfl-fastssim.png http://people.xiph.org/~unlord/head_vs_cfl-luma4x4-psnr.png http://people.xiph.org/~unlord/head_vs_cfl-luma4x4-psnrhvs.png http://people.xiph.org/~unlord/head_vs_cfl-luma4x4-ssim.png http://people.xiph.org/~unlord/head_vs_cfl-luma4x4-fastssim.png http://people.xiph.org/~unlord/head_vs_cfl-luma8x8-psnr.png http://people.xiph.org/~unlord/head_vs_cfl-luma8x8-psnrhvs.png http://people.xiph.org/~unlord/head_vs_cfl-luma8x8-ssim.png http://people.xiph.org/~unlord/head_vs_cfl-luma8x8-fastssim.png http://people.xiph.org/~unlord/head_vs_cfl-luma16x16-psnr.png http://people.xiph.org/~unlord/head_vs_cfl-luma16x16-psnrhvs.png http://people.xiph.org/~unlord/head_vs_cfl-luma16x16-ssim.png http://people.xiph.org/~unlord/head_vs_cfl-luma16x16-fastssim.png Ran cfl_stats on subset1 and subset3, Cb predicts well, Cr not so much: https://pastebin.mozilla.org/3586503 In CFL fix the 4x4 case doesn't improve very much compared to the 8x8 and 16x16 case, but this isn't entirely surprising because the fix's effect was not as substantial in this case. We thought the Cr/Cb difference sounded weird, might be explained by content, suggested swapping the two planes on input. Jm replaced the PVQ encoder code in the encoder, but did not replace the code in the decoder as there is no decoder written yet. So PVQ no longer works in the codebase right now. The current encoder code is 8x8 only as well, so it will us scalar for other blocksizes. [We lapsed into Opus discussion at this point, targeting 1.1-rc release perhaps next monday]