2015年JCTVC-G399.pptVIP

Comparison of Compression Performance of HEVC Working Draft 4 with AVC High Profile Bin Li, Gary J. Sullivan, Jizheng Xu University of Science and Technology of ChinaMicrosoft Corp. Purpose Test condition HEVC WD4 Default HE anchor of HM-4.0 AVC High Profile Modified JM 18.0 by encoder only actions 8 bit encoding/decoding (High Profile) Weighted prediction disabled. (HM-4.0 does not support weighted prediction, and if supported the impact should be similar) Negligible impact for B frames 1% average benefit for LD P cases Outline Generate a better AVC anchor Compare JM with HM Some more information on bit depth A better AVC anchor Try different coding structure on AVC to find a better AVC anchor 1+3 references for low-delay cases B-frame coding at temporal level 0 for random-access cases For All Intra case, nothing special needs to be configured JM Anchor of Random access HM coding order Use 0 8 4 2 6 1 3 5 7, rather than 0 8 4 2 1 3 6 5 7 Average difference 0.0% in JM Increase number of reference frames to 8 to match the HM referencing structure 8 reference frames are used, although 6 is enough HM buffers at most 10 reference frames P or B frames at temporal 0? (to be tested) If B frames, list0 == list1? (to be tested) JM RA Anchor Anchor is P at temporal level 0 Better anchor is B at temporal level 0 and list1=list0   B at T0 B at T0, l1=l0 Class A −1.2% −1.5% Class B −1.5% −1.9% Class C −0.2% −0.4% Class D +0.3% +0.0% Average −0.7% −1.0% JM anchor of Low delay Use the same QP setting as HM RateGopSize = 4 4+0 referencing structure or 1+3 referencing structure? (to be tested) B or P frames? (to be tested) If B frames, list0 == list1? (to be tested) JM LD Anchor Anchor is P frames, 4+0 referencing structure Better anchor is P frames, 1+3 referencing structure   P,1+3 B,1+3 B,1+3 l1=l0 Class B −2.0% −1.6% −1.2% Class C −3.4% −4.4% −4.8% Class D −3.3% −4.3% −3.8% Class E −1.2% +3.8% +3.3% Average −2.5% −2.0% −1.9% Our JM Anchor VS. D181 JM Anchor Nebuta and SteamLo