shithub: libvpx

Info  •  Files  •  Log  •  Branches

Improve two pass VBR accuracy.

Adjustments to the GF interval choice and minimum boost.
Adjustment to the calculation of 2 pass worst q.
Compared to 09/29 head there is metrics hit on derf of
(-0.123%,-0.191%)

Compared to the September 29 head and a baseline on
September 18 baseline the accuracy of the VBR rate control
measured on the derf set is as follows:-

Mean error %  / Mean abs(error %)
Sept 18 baseline (-7.0% / 14.76%)
Sept 29 head (-15.7%, 19.8%)
This check in (-1.5% / 14.4%)

The mean undershoot is reduced slightly but the
worst case overshoot on e.g. harbour/highway is
increased. This will be addressed in a later patch.

Change-Id: Iffd9b0ab7432a131c98fbaaa82d1e5b40be72b58

--- a/vp9/encoder/vp9_encodeframe.c

+++ b/vp9/encoder/vp9_encodeframe.c

@@ -1758,7 +1758,7 @@

   // We must have chosen a partitioning and encoding or we'll fail later on.

   // No other opportunities for success.

-  if ( bsize == BLOCK_64X64)

+  if (bsize == BLOCK_64X64)

     assert(chosen_rate < INT_MAX && chosen_dist < INT64_MAX);

   if (do_recon) {

--- a/vp9/encoder/vp9_firstpass.c

+++ b/vp9/encoder/vp9_firstpass.c

@@ -39,9 +39,9 @@

 #define ARF_STATS_OUTPUT    0

 #define BOOST_FACTOR        12.5

-#define ERR_DIVISOR         100.0

-#define FACTOR_PT_LOW       0.5

-#define FACTOR_PT_HIGH      0.9

+#define ERR_DIVISOR         125.0

+#define FACTOR_PT_LOW       0.70

+#define FACTOR_PT_HIGH      0.90

 #define FIRST_PASS_Q        10.0

 #define GF_MAX_BOOST        96.0

 #define INTRA_MODE_PENALTY  1024

@@ -1057,7 +1057,7 @@

   // Adjustment based on actual quantizer to power term.

   const double power_term =

-      MIN(vp9_convert_qindex_to_q(q, bit_depth) * 0.0125 + pt_low, pt_high);

+      MIN(vp9_convert_qindex_to_q(q, bit_depth) * 0.01 + pt_low, pt_high);

   // Calculate correction factor.

   if (power_term < 1.0)

@@ -1066,6 +1066,11 @@

   return fclamp(pow(error_term, power_term), 0.05, 5.0);

 }

+// Larger image formats are expected to be a little harder to code relatively

+// given the same prediction error score. This in part at least relates to the

+// increased size and hence coding cost of motion vectors.

+#define EDIV_SIZE_FACTOR 800

+

 static int get_twopass_worst_quality(const VP9_COMP *cpi,

                                      const FIRSTPASS_STATS *stats,

                                      int section_target_bandwidth) {

@@ -1079,8 +1084,10 @@

     const double section_err = stats->coded_error / stats->count;

     const double err_per_mb = section_err / num_mbs;

     const double speed_term = 1.0 + 0.04 * oxcf->speed;

+    const double ediv_size_correction = num_mbs / EDIV_SIZE_FACTOR;

     const int target_norm_bits_per_mb = ((uint64_t)section_target_bandwidth <<

                                          BPER_MB_NORMBITS) / num_mbs;

+

     int q;

     int is_svc_upper_layer = 0;

     if (is_two_pass_svc(cpi) && cpi->svc.spatial_layer_id > 0)

@@ -1090,7 +1097,7 @@

     // content at the given rate.

     for (q = rc->best_quality; q < rc->worst_quality; ++q) {

       const double factor =

-          calc_correction_factor(err_per_mb, ERR_DIVISOR,

+          calc_correction_factor(err_per_mb, ERR_DIVISOR - ediv_size_correction,

                                  is_svc_upper_layer ? SVC_FACTOR_PT_LOW :

                                  FACTOR_PT_LOW, FACTOR_PT_HIGH, q,

                                  cpi->common.bit_depth);

@@ -1735,7 +1742,7 @@

       // bits to spare and are better with a smaller interval and smaller boost.

       // At high Q when there are few bits to spare we are better with a longer

       // interval to spread the cost of the GF.

-      active_max_gf_interval = 12 + MIN(4, (int_max_q / 32));

+      active_max_gf_interval = 12 + MIN(4, (int_max_q / 24));

       if (active_max_gf_interval > rc->max_gf_interval)

         active_max_gf_interval = rc->max_gf_interval;

     }

--- a/vp9/encoder/vp9_ratectrl.c

+++ b/vp9/encoder/vp9_ratectrl.c

@@ -177,6 +177,9 @@

   const double q = vp9_convert_qindex_to_q(qindex, bit_depth);

   int enumerator = frame_type == KEY_FRAME ? 2700000 : 1800000;

+  assert(correction_factor <= MAX_BPB_FACTOR &&

+         correction_factor >= MIN_BPB_FACTOR);

+

   // q based adjustment to baseline enumerator

   enumerator += (int)(enumerator * q) >> 12;

   return (int)(enumerator * correction_factor / q);

@@ -187,7 +190,8 @@

                               vpx_bit_depth_t bit_depth) {

   const int bpm = (int)(vp9_rc_bits_per_mb(frame_type, q, correction_factor,

                                            bit_depth));

-  return ((uint64_t)bpm * mbs) >> BPER_MB_NORMBITS;

+  return MAX(FRAME_OVERHEAD_BITS,

+             (int)((uint64_t)bpm * mbs) >> BPER_MB_NORMBITS);

 }

 int vp9_rc_clamp_pframe_target_size(const VP9_COMP *const cpi, int target) {

@@ -410,7 +414,7 @@

                                                  rate_correction_factor,

                                                  cm->bit_depth);

   // Work out a size correction factor.

-  if (projected_size_based_on_q > 0)

+  if (projected_size_based_on_q > FRAME_OVERHEAD_BITS)

     correction_factor = (100 * cpi->rc.projected_frame_size) /

                             projected_size_based_on_q;