From 31228b3595d17e8323aede7bb364bd731c192820 Mon Sep 17 00:00:00 2001 From: Paul Wilkins Date: Wed, 20 Jun 2018 17:21:49 +0100 Subject: [PATCH] Improved coding on slide show content. This patch adds in detection of slide show content and allows for coding of long GF only groups up to a length of 240 frames rather than coding a large number of shorter ARF groups that gradually lower the Q. In test samples this patch gave rise to a substantial improvement in overall psnr and a drop in data rate. In some cases the average psnr fell, however, with the boost and minQ values set as they are. This is to be expected because average psnr is dominated by the best frames in the sequence and previously a relatively poor key frame could be followed by progressively better alt refs. For example a key frame at q7.5 but subsequent alt refs improving it to lossless. For slides displayed for several seconds, savings of >= 20% (or commensurate quality gains) are likely. This patch allows for long GF groups in static sections before and after complex transitions (e.g. fades) with one or more normal ARF groups during the transition. However, it enforces a single "normal" length GF group after the transition before any extended group is allowed. The reason for this is that the ARF that spans the transition my not have a very high quality and hence may not be a good GF for the long static section that follows. Change-Id: I66cc404c3b85e87dae9829b49d9d631cbf04e037 --- vp9/encoder/vp9_firstpass.c | 64 +++++++++++++++++++++++++++++++++------------ vp9/encoder/vp9_firstpass.h | 16 ++++++------ vp9/encoder/vp9_ratectrl.c | 17 +++++++----- vp9/encoder/vp9_ratectrl.h | 8 ++++++ 4 files changed, 73 insertions(+), 32 deletions(-) diff --git a/vp9/encoder/vp9_firstpass.c b/vp9/encoder/vp9_firstpass.c index c8cf37e..b7456c1 100644 --- a/vp9/encoder/vp9_firstpass.c +++ b/vp9/encoder/vp9_firstpass.c @@ -2397,7 +2397,8 @@ static void define_gf_group_structure(VP9_COMP *cpi) { // Define middle frame mid_frame_idx = frame_index + (rc->baseline_gf_interval >> 1) - 1; - normal_frames = (rc->baseline_gf_interval - rc->source_alt_ref_pending); + normal_frames = + rc->baseline_gf_interval - (key_frame || rc->source_alt_ref_pending); for (i = 0; i < normal_frames; ++i) { int arf_idx = 0; if (twopass->stats_in >= twopass->stats_in_end) break; @@ -2859,8 +2860,9 @@ static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) { loop_decay_rate = get_prediction_decay_rate(cpi, &next_frame); // Monitor for static sections. - zero_motion_accumulator = VPXMIN( - zero_motion_accumulator, get_zero_motion_factor(cpi, &next_frame)); + if ((rc->frames_since_key + i - 1) > 1) { + zero_motion_accumulator *= get_zero_motion_factor(cpi, &next_frame); + } // Break clause to detect very still sections after motion. For example, // a static image after a fade or other transition. @@ -2881,9 +2883,18 @@ static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) { } // Break out conditions. - if ( - // Break at active_max_gf_interval unless almost totally static. - ((i >= active_max_gf_interval) && (zero_motion_accumulator < 0.995)) || + // Break at maximum of active_max_gf_interval unless almost totally static. + // + // Note that the addition of a test of rc->source_alt_ref_active is + // deliberate. The effect of this is that after a normal altref group even + // if the material is static there will be one normal length GF group + // before allowing longer GF groups. The reason for this is that in cases + // such as slide shows where slides are separated by a complex transition + // such as a fade, the arf group spanning the transition may not be coded + // at a very high quality and hence this frame (with its overlay) is a + // poor golden frame to use for an extended group. + if (((i >= active_max_gf_interval) && + ((zero_motion_accumulator < 0.995) || (rc->source_alt_ref_active))) || ( // Don't break out with a very short interval. (i >= active_min_gf_interval) && @@ -2903,8 +2914,8 @@ static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) { rc->constrained_gf_group = (i >= rc->frames_to_key) ? 1 : 0; // Should we use the alternate reference frame. - if (allow_alt_ref && (i < cpi->oxcf.lag_in_frames) && - (i >= rc->min_gf_interval)) { + if ((zero_motion_accumulator < 0.995) && allow_alt_ref && + (i < cpi->oxcf.lag_in_frames) && (i >= rc->min_gf_interval)) { const int forward_frames = (rc->frames_to_key - i >= i - 1) ? i - 1 : VPXMAX(0, rc->frames_to_key - i); @@ -2931,8 +2942,11 @@ static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) { rc->gfu_boost = VPXMIN((int)rc->gfu_boost, i * 200); #endif - // Set the interval until the next gf. - rc->baseline_gf_interval = i - (is_key_frame || rc->source_alt_ref_pending); + rc->baseline_gf_interval = + ((twopass->kf_zeromotion_pct >= STATIC_KF_GROUP_THRESH) && + (i >= rc->frames_to_key)) + ? i + : (i - (is_key_frame || rc->source_alt_ref_pending)); rc->frames_till_gf_update_due = rc->baseline_gf_interval; @@ -3216,6 +3230,7 @@ static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) { vp9_zero(next_frame); cpi->common.frame_type = KEY_FRAME; + rc->frames_since_key = 0; // Reset the GF group data structures. vp9_zero(*gf_group); @@ -3360,13 +3375,22 @@ static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) { for (i = 0; i < (rc->frames_to_key - 1); ++i) { if (EOF == input_stats(twopass, &next_frame)) break; - if (i <= KF_BOOST_SCAN_MAX_FRAMES) { + // The zero motion test here insures that if we mark a kf group as static + // it is static throughout not just the first KF_BOOST_SCAN_MAX_FRAMES. + // It also allows for a larger boost on long static groups. + if ((i <= KF_BOOST_SCAN_MAX_FRAMES) || (zero_motion_accumulator >= 0.99)) { double frame_boost; double zm_factor; // Monitor for static sections. - zero_motion_accumulator = VPXMIN( - zero_motion_accumulator, get_zero_motion_factor(cpi, &next_frame)); + // First frame in kf group the second ref indicator is invalid. + if (i > 0) { + zero_motion_accumulator = VPXMIN( + zero_motion_accumulator, get_zero_motion_factor(cpi, &next_frame)); + } else { + zero_motion_accumulator = + next_frame.pcnt_inter - next_frame.pcnt_motion; + } // Factor 0.75-1.25 based on how much of frame is static. zm_factor = (0.75 + (zero_motion_accumulator / 2.0)); @@ -3403,10 +3427,16 @@ static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) { twopass->section_intra_rating = calculate_section_intra_ratio( start_position, twopass->stats_in_end, rc->frames_to_key); - // Apply various clamps for min and max boost - rc->kf_boost = VPXMAX((int)boost_score, (rc->frames_to_key * 3)); - rc->kf_boost = VPXMAX(rc->kf_boost, MIN_KF_TOT_BOOST); - rc->kf_boost = VPXMIN(rc->kf_boost, MAX_KF_TOT_BOOST); + // Special case for static / slide show content but dont apply + // if the kf group is very short. + if ((zero_motion_accumulator > 0.99) && (rc->frames_to_key > 8)) { + rc->kf_boost = VPXMAX((rc->frames_to_key * 100), MAX_KF_TOT_BOOST); + } else { + // Apply various clamps for min and max boost + rc->kf_boost = VPXMAX((int)boost_score, (rc->frames_to_key * 3)); + rc->kf_boost = VPXMAX(rc->kf_boost, MIN_KF_TOT_BOOST); + rc->kf_boost = VPXMIN(rc->kf_boost, MAX_KF_TOT_BOOST); + } // Work out how many bits to allocate for the key frame itself. kf_bits = calculate_boost_bits((rc->frames_to_key - 1), rc->kf_boost, diff --git a/vp9/encoder/vp9_firstpass.h b/vp9/encoder/vp9_firstpass.h index 404175d..2717913 100644 --- a/vp9/encoder/vp9_firstpass.h +++ b/vp9/encoder/vp9_firstpass.h @@ -134,14 +134,14 @@ typedef enum { typedef struct { unsigned char index; unsigned char first_inter_index; - RATE_FACTOR_LEVEL rf_level[(MAX_LAG_BUFFERS * 2) + 1]; - FRAME_UPDATE_TYPE update_type[(MAX_LAG_BUFFERS * 2) + 1]; - unsigned char arf_src_offset[(MAX_LAG_BUFFERS * 2) + 1]; - unsigned char arf_update_idx[(MAX_LAG_BUFFERS * 2) + 1]; - unsigned char arf_ref_idx[(MAX_LAG_BUFFERS * 2) + 1]; - unsigned char brf_src_offset[(MAX_LAG_BUFFERS * 2) + 1]; - unsigned char bidir_pred_enabled[(MAX_LAG_BUFFERS * 2) + 1]; - int bit_allocation[(MAX_LAG_BUFFERS * 2) + 1]; + RATE_FACTOR_LEVEL rf_level[MAX_STATIC_GF_GROUP_LENGTH + 2]; + FRAME_UPDATE_TYPE update_type[MAX_STATIC_GF_GROUP_LENGTH + 2]; + unsigned char arf_src_offset[MAX_STATIC_GF_GROUP_LENGTH + 2]; + unsigned char arf_update_idx[MAX_STATIC_GF_GROUP_LENGTH + 2]; + unsigned char arf_ref_idx[MAX_STATIC_GF_GROUP_LENGTH + 2]; + unsigned char brf_src_offset[MAX_STATIC_GF_GROUP_LENGTH + 2]; + unsigned char bidir_pred_enabled[MAX_STATIC_GF_GROUP_LENGTH + 2]; + int bit_allocation[MAX_STATIC_GF_GROUP_LENGTH + 2]; } GF_GROUP; typedef struct { diff --git a/vp9/encoder/vp9_ratectrl.c b/vp9/encoder/vp9_ratectrl.c index f5a0641..22ae726 100644 --- a/vp9/encoder/vp9_ratectrl.c +++ b/vp9/encoder/vp9_ratectrl.c @@ -1242,6 +1242,14 @@ static int rc_pick_q_and_bounds_two_pass(const VP9_COMP *cpi, int *bottom_index, // Baseline value derived from cpi->active_worst_quality and kf boost. active_best_quality = get_kf_active_quality(rc, active_worst_quality, cm->bit_depth); + if (cpi->twopass.kf_zeromotion_pct >= STATIC_KF_GROUP_THRESH) { + active_best_quality /= 4; + } + + // Dont allow the active min to be lossless (q0) unlesss the max q + // already indicates lossless. + active_best_quality = + VPXMIN(active_worst_quality, VPXMAX(1, active_best_quality)); // Allow somewhat lower kf minq with small image formats. if ((cm->width * cm->height) <= (352 * 288)) { @@ -2184,13 +2192,8 @@ void vp9_rc_set_gf_interval_range(const VP9_COMP *const cpi, rc->max_gf_interval = vp9_rc_get_default_max_gf_interval( cpi->framerate, rc->min_gf_interval); - // Extended interval for genuinely static scenes - rc->static_scene_max_gf_interval = MAX_LAG_BUFFERS * 2; - - if (is_altref_enabled(cpi)) { - if (rc->static_scene_max_gf_interval > oxcf->lag_in_frames - 1) - rc->static_scene_max_gf_interval = oxcf->lag_in_frames - 1; - } + // Extended max interval for genuinely static scenes like slide shows. + rc->static_scene_max_gf_interval = MAX_STATIC_GF_GROUP_LENGTH; if (rc->max_gf_interval > rc->static_scene_max_gf_interval) rc->max_gf_interval = rc->static_scene_max_gf_interval; diff --git a/vp9/encoder/vp9_ratectrl.h b/vp9/encoder/vp9_ratectrl.h index 02e5783..373c6a3 100644 --- a/vp9/encoder/vp9_ratectrl.h +++ b/vp9/encoder/vp9_ratectrl.h @@ -34,6 +34,14 @@ extern "C" { #define FRAME_OVERHEAD_BITS 200 +// Threshold used to define a KF group as static (e.g. a slide show). +// Essentially this means that no frame in the group has more than 1% of MBs +// that are not marked as coded with 0,0 motion in the first pass. +#define STATIC_KF_GROUP_THRESH 99 + +// The maximum duration of a GF group that is static (for example a slide show). +#define MAX_STATIC_GF_GROUP_LENGTH 250 + typedef enum { INTER_NORMAL = 0, INTER_HIGH = 1, -- 2.7.4