Figure 10.
The panels compare H M 16.26, V T M 13.2, D V C, D V C pro, D C V C, C A N F-V C, D H V C, V C T and A T E M-Hybrid C A. The horizontal axis is b p p, and the vertical axis is P S N R. For H E V C-B, b p p spans about 0.05 to 0.20 and P S N R spans 32 to 36. A T E M-Hybrid C A rises from about 33.7 at 0.05 b p p to about 35.7 near 0.20 b p p. V T M 13.2 reaches about 35.9 near 0.18 b p p. H M 16.26 and V C T approach about 35.7 and 35.5, respectively, at higher b p p values. The remaining learned methods rise from roughly 32 to 33 at lower b p p values to about 34.5 to 35.2 at higher values. For H E V C-C, b p p spans about 0.08 to 0.40 and P S N R spans 30 to 35. A T E M-Hybrid C A increases from about 32.4 near 0.10 b p p to about 34.2 near 0.40 b p p. V T M 13.2 reaches about 35.0 near 0.28 b p p. H M 16.26 rises to about 34.9 near 0.37 b p p. V C T rises from about 32.3 to about 33.9. D V C, D V C pro, D C V C and D H V C remain between roughly 30 and 32.3 over their plotted ranges. For H E V C-D, b p p spans about 0.10 to 0.50 and P S N R spans 30 to 35. A T E M-Hybrid C A rises from about 32.1 near 0.11 b p p to about 34.3 near 0.41 b p p. V T M 13.2 reaches approximately 35 near 0.28 b p p. H M 16.26 approaches about 35 near 0.38 b p p. V C T rises from about 31.9 to about 34.0. D V C, D V C pro, D C V C and D H V C occupy lower P S N R ranges, reaching approximately 32.0, 32.3, 32.7 and 33.5, respectively. For H E V C-E, b p p spans about 0.01 to 0.10 and P S N R spans 35 to 40. A T E M-Hybrid C A increases from about 36.2 near 0.017 b p p to about 39.6 near 0.096 b p p. V T M 13.2 reaches around 40 near 0.04 b p p. H M 16.26 reaches about 39.8 near 0.064 b p p. V C T rises from about 36.0 to above 39 near 0.10 b p p. D V C, D V C pro, D C V C, C A N F-V C and D H V C rise through approximately 35 to 38.9 across their plotted b p p ranges.Rate-distortion curves comparing the VCT baseline against the best performing Conv-Attn Hybrid Mixer in ATEM on the HEVC Class B, C, D and E data sets. Traditional codecs (HM and VTM) are indicated by dashed lines for reference. For broader context, several pioneering LVC methods are also plotted, including DVC (Lu et al., 2018), DVC-Pro (Lu et al., 2020), DCVC (Li et al., 2021), DHVC (Lu et al., 2023), and CANF-VC (Ho et al., 2022)
Source: Authors’ own work