shithub: libvpx

Info  •  Files  •  Log  •  Branches

Add intra complexity and brightness weight to first pass.

The aim of this patch is to apply a positive weighting to
frames that have a significant number of blocks that are
of low spatial complexity and are dark. The rationale behind
this is that artifacts tend to be more visible in such frames.

In this patch the weight is only applied in regard to the distribution
of bits between frames. Hence if all the frames share similar
characteristics (as is the case for most of our short test clips) there
will be little or no net effect.

However, the effect can be seen on some longer form test content.

For example Tears of steel baseline test:
2323.09 Kbit/s opsnr 39.915 ssim 74.729
With this patch:-
2213.34 Kbit/s opsnr 39.963 ssim 74.808
(Sligtly better metrics and about 5% smaller)

The weighting may well need some further tuning along side changes
to the aq modes.

Change-Id: Ieced379bca03938166ab87b2b97f55d94948904c

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

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

@@ -53,6 +53,7 @@

 #define MIN_KF_BOOST        300

 #define NEW_MV_MODE_PENALTY 32

 #define SVC_FACTOR_PT_LOW   0.45

+#define DARK_THRESH         64

 #define DOUBLE_DIVIDE_CHECK(x) ((x) < 0 ? (x) - 0.000001 : (x) + 0.000001)

@@ -106,10 +107,11 @@

     FILE *fpfile;

     fpfile = fopen("firstpass.stt", "a");

-    fprintf(fpfile, "%12.0f %12.0f %12.0f %12.0f %12.4f %12.4f"

+    fprintf(fpfile, "%12.0f %12.4f %12.0f %12.0f %12.0f %12.4f %12.4f"

             "%12.4f %12.4f %12.4f %12.4f %12.4f %12.4f %12.4f"

             "%12.0f %12.0f %12.4f %12.0f %12.0f %12.4f\n",

             stats->frame,

+            stats->weight,

             stats->intra_error,

             stats->coded_error,

             stats->sr_coded_error,

@@ -144,7 +146,8 @@

 #endif

 static void zero_stats(FIRSTPASS_STATS *section) {

-  section->frame      = 0.0;

+  section->frame = 0.0;

+  section->weight = 0.0;

   section->intra_error = 0.0;

   section->coded_error = 0.0;

   section->sr_coded_error = 0.0;

@@ -168,6 +171,7 @@

 static void accumulate_stats(FIRSTPASS_STATS *section,

                              const FIRSTPASS_STATS *frame) {

   section->frame += frame->frame;

+  section->weight += frame->weight;

   section->spatial_layer_id = frame->spatial_layer_id;

   section->intra_error += frame->intra_error;

   section->coded_error += frame->coded_error;

@@ -191,6 +195,7 @@

 static void subtract_stats(FIRSTPASS_STATS *section,

                            const FIRSTPASS_STATS *frame) {

   section->frame -= frame->frame;

+  section->weight -= frame->weight;

   section->intra_error -= frame->intra_error;

   section->coded_error -= frame->coded_error;

   section->sr_coded_error -= frame->sr_coded_error;

@@ -217,10 +222,11 @@

                                      const VP9EncoderConfig *oxcf,

                                      const FIRSTPASS_STATS *this_frame) {

   const FIRSTPASS_STATS *const stats = &twopass->total_stats;

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

-  const double modified_error = av_err *

-      pow(this_frame->coded_error / DOUBLE_DIVIDE_CHECK(av_err),

-          oxcf->two_pass_vbrbias / 100.0);

+  const double av_weight = stats->weight / stats->count;

+  const double av_err = (stats->coded_error * av_weight) / stats->count;

+  const double modified_error =

+    av_err * pow(this_frame->coded_error * this_frame->weight /

+                 DOUBLE_DIVIDE_CHECK(av_err), oxcf->two_pass_vbrbias / 100.0);

   return fclamp(modified_error,

                 twopass->modified_error_min, twopass->modified_error_max);

 }

@@ -480,6 +486,8 @@

   const YV12_BUFFER_CONFIG *first_ref_buf = lst_yv12;

   LAYER_CONTEXT *const lc = is_two_pass_svc(cpi) ?

         &cpi->svc.layer_context[cpi->svc.spatial_layer_id] : NULL;

+  double intra_factor = 0.0;

+  double brightness_factor = 0.0;

 #if CONFIG_FP_MB_STATS

   if (cpi->use_fp_mb_stats) {

@@ -587,6 +595,9 @@

       const int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);

       double error_weight = 1.0;

       const BLOCK_SIZE bsize = get_bsize(cm, mb_row, mb_col);

+      double log_intra;

+      int level_sample;

+

 #if CONFIG_FP_MB_STATS

       const int mb_index = mb_row * cm->mb_cols + mb_col;

 #endif

@@ -635,6 +646,25 @@

       }

 #endif  // CONFIG_VP9_HIGHBITDEPTH

+      log_intra = log(this_error + 1.0);

+      if (log_intra < 10.0)

+        intra_factor += 1.0 + ((10.0 - log_intra) * 0.05);

+      else

+        intra_factor += 1.0;

+

+#if CONFIG_VP9_HIGHBITDEPTH

+      if (cm->use_highbitdepth)

+        level_sample = CONVERT_TO_SHORTPTR(x->plane[0].src.buf)[0];

+      else

+        level_sample = x->plane[0].src.buf[0];

+#else

+      level_sample = x->plane[0].src.buf[0];

+#endif

+      if ((level_sample < DARK_THRESH) && (log_intra < 9.0))

+        brightness_factor += 1.0 + (0.01 * (DARK_THRESH - level_sample));

+      else

+        brightness_factor += 1.0;

+

       if (cpi->oxcf.aq_mode == VARIANCE_AQ) {

         vp9_clear_system_state();

         this_error = (int)(this_error * error_weight);

@@ -943,6 +973,10 @@

         cpi->oxcf.allow_spatial_resampling ? cpi->initial_mbs : cpi->common.MBs;

     const double min_err = 200 * sqrt(num_mbs);

+    intra_factor = intra_factor / (double)num_mbs;

+    brightness_factor = brightness_factor / (double)num_mbs;

+    fps.weight = intra_factor * brightness_factor;

+

     fps.frame = cm->current_video_frame;

     fps.spatial_layer_id = cpi->svc.spatial_layer_id;

     fps.coded_error = (double)(coded_error >> 8) + min_err;

@@ -1094,7 +1128,8 @@

     // 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 - ediv_size_correction,

+          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);

@@ -1319,7 +1354,7 @@

   const double lq =

     vp9_convert_qindex_to_q(cpi->rc.avg_frame_qindex[INTER_FRAME],

                             cpi->common.bit_depth);

-  const double boost_correction = MIN((0.5 + (lq * 0.015)), 1.5);

+  const double boost_q_correction = MIN((0.5 + (lq * 0.015)), 1.5);

   const int num_mbs =

       cpi->oxcf.allow_spatial_resampling ? cpi->initial_mbs : cpi->common.MBs;

@@ -1326,7 +1361,7 @@

   // Underlying boost factor is based on inter error ratio.

   frame_boost = (BASELINE_ERR_PER_MB * num_mbs) /

                 DOUBLE_DIVIDE_CHECK(this_frame->coded_error);

-  frame_boost = frame_boost * BOOST_FACTOR * boost_correction;

+  frame_boost = frame_boost * BOOST_FACTOR * boost_q_correction;

   // Increase boost for frames where new data coming into frame (e.g. zoom out).

   // Slightly reduce boost if there is a net balance of motion out of the frame

@@ -1337,7 +1372,7 @@

   else

     frame_boost += frame_boost * (this_frame_mv_in_out / 2.0);

-  return MIN(frame_boost, max_boost * boost_correction);

+  return MIN(frame_boost, max_boost * boost_q_correction);

 }

 static int calc_arf_boost(VP9_COMP *cpi, int offset,

--- a/vp9/encoder/vp9_firstpass.h

+++ b/vp9/encoder/vp9_firstpass.h

@@ -43,6 +43,7 @@

 typedef struct {

   double frame;

+  double weight;

   double intra_error;

   double coded_error;

   double sr_coded_error;