Small change to av1_frame_type_qdelta()
This change gives slight improvements to PSNR and VMAF
metrics with a 4 layer ARF structure and max interval of 16
but slightly larger gains when 5 levels are used when the max
interval is increased to 32. Slightly worse in SSIM for the
low and mid res sets.
STATS_CHANGED
(av psnr, ov psnr, ssim, psnr-hvs, VMAF)
Max ARF interval 16
Low res: 0.008 -0.017 0.171 -0.053 -0.088
UGC-360: -0.014 -0.156 0.035 -0.105 -0.204
Mid res: -0.005 -0.054 0.237 -0.114 -0.222
Max ARF interval 32
Low res: -0.069 -0.127 0.155 -0.123 -0.226
UGC-360: -0.118 -0.374 -0.106 -0.280 -0.425
Mid res: -0.090 -0.157 0.272 -0.217 -0.323
Change-Id: I5c991f193218b8a19738e94ee8feab71cc5dc2bd
diff --git a/av1/encoder/ratectrl.c b/av1/encoder/ratectrl.c
index 7a895ca..a5ef72d 100644
--- a/av1/encoder/ratectrl.c
+++ b/av1/encoder/ratectrl.c
@@ -1100,7 +1100,7 @@
rate_factor = rate_factor_deltas[rf_lvl];
if (rf_lvl == GF_ARF_LOW) {
- rate_factor -= (cpi->gf_group.layer_depth[cpi->gf_group.index] - 2) * 0.2;
+ rate_factor -= (cpi->gf_group.layer_depth[cpi->gf_group.index] - 2) * 0.1;
rate_factor = AOMMAX(rate_factor, 1.0);
}
return av1_compute_qdelta_by_rate(&cpi->rc, frame_type, q, rate_factor,
