Git Inbox Mirror of the ffmpeg-devel mailing list - see https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
 help / color / mirror / Atom feed
From: "Clément Bœsch" <u@pkh.me>
To: ffmpeg-devel@ffmpeg.org
Cc: "Clément Bœsch" <u@pkh.me>
Subject: [FFmpeg-devel] [PATCH v2 11/32] avfilter/palettegen: define the best axis to cut using the squared error
Date: Wed, 28 Dec 2022 00:17:53 +0100
Message-ID: <20221227231814.2520181-12-u@pkh.me> (raw)
In-Reply-To: <20221227231814.2520181-1-u@pkh.me>

This is following the results from personal research¹.

¹: https://github.com/ubitux/research/tree/main/color-quantization#results
---
 libavfilter/vf_palettegen.c        | 42 ++++++++++++++++++------------
 tests/ref/fate/filter-palettegen-1 |  2 +-
 tests/ref/fate/filter-palettegen-2 |  2 +-
 3 files changed, 27 insertions(+), 19 deletions(-)

diff --git a/libavfilter/vf_palettegen.c b/libavfilter/vf_palettegen.c
index a047c75599..ed1448755c 100644
--- a/libavfilter/vf_palettegen.c
+++ b/libavfilter/vf_palettegen.c
@@ -147,31 +147,39 @@ static av_always_inline int diff(const uint32_t a, const uint32_t b)
 
 static void compute_box_stats(PaletteGenContext *s, struct range_box *box)
 {
-    int rr, gr, br;
+    int avg[3];
+    int64_t er2[3] = {0};
 
-    /* compute the box weight (sum all the weights of the colors in the
-     * range) and its boundings */
-    uint8_t min[3] = {0xff, 0xff, 0xff};
-    uint8_t max[3] = {0x00, 0x00, 0x00};
+    /* Compute average color */
+    uint64_t sr = 0, sg = 0, sb = 0;
     box->weight = 0;
     for (int i = box->start; i < box->start + box->len; i++) {
         const struct color_ref *ref = s->refs[i];
-        const uint32_t rgb = ref->color;
-        const uint8_t r = rgb >> 16 & 0xff, g = rgb >> 8 & 0xff, b = rgb & 0xff;
-        min[0] = FFMIN(r, min[0]), max[0] = FFMAX(r, max[0]);
-        min[1] = FFMIN(g, min[1]), max[1] = FFMAX(g, max[1]);
-        min[2] = FFMIN(b, min[2]), max[2] = FFMAX(b, max[2]);
+        sr += (ref->color >> 16 & 0xff) * ref->count;
+        sg += (ref->color >>  8 & 0xff) * ref->count;
+        sb += (ref->color       & 0xff) * ref->count;
         box->weight += ref->count;
     }
+    avg[0] = sr / box->weight;
+    avg[1] = sg / box->weight;
+    avg[2] = sb / box->weight;
 
-    /* define the axis to sort by according to the widest range of colors */
-    rr = max[0] - min[0];
-    gr = max[1] - min[1];
-    br = max[2] - min[2];
+    /* Compute squared error of each color channel */
+    for (int i = box->start; i < box->start + box->len; i++) {
+        const struct color_ref *ref = s->refs[i];
+        const int64_t dr = (int)(ref->color >> 16 & 0xff) - avg[0];
+        const int64_t dg = (int)(ref->color >>  8 & 0xff) - avg[1];
+        const int64_t db = (int)(ref->color       & 0xff) - avg[2];
+        er2[0] += dr * dr * ref->count;
+        er2[1] += dg * dg * ref->count;
+        er2[2] += db * db * ref->count;
+    }
+
+    /* Define the best axis candidate for cutting the box */
     box->major_axis = 1; // pick green by default (the color the eye is the most sensitive to)
-    if (br >= rr && br >= gr) box->major_axis = 2;
-    if (rr >= gr && rr >= br) box->major_axis = 0;
-    if (gr >= rr && gr >= br) box->major_axis = 1; // prefer green again
+    if (er2[2] >= er2[0] && er2[2] >= er2[1]) box->major_axis = 2;
+    if (er2[0] >= er2[1] && er2[0] >= er2[2]) box->major_axis = 0;
+    if (er2[1] >= er2[0] && er2[1] >= er2[2]) box->major_axis = 1; // prefer green again
 }
 
 /**
diff --git a/tests/ref/fate/filter-palettegen-1 b/tests/ref/fate/filter-palettegen-1
index bebfd24e19..278d831846 100644
--- a/tests/ref/fate/filter-palettegen-1
+++ b/tests/ref/fate/filter-palettegen-1
@@ -3,4 +3,4 @@
 #codec_id 0: rawvideo
 #dimensions 0: 16x16
 #sar 0: 1/1
-0,          0,          0,        1,     1024, 0x3395ef5a
+0,          0,          0,        1,     1024, 0x394ee723
diff --git a/tests/ref/fate/filter-palettegen-2 b/tests/ref/fate/filter-palettegen-2
index 9abec0fe8e..e9bc635c81 100644
--- a/tests/ref/fate/filter-palettegen-2
+++ b/tests/ref/fate/filter-palettegen-2
@@ -3,4 +3,4 @@
 #codec_id 0: rawvideo
 #dimensions 0: 16x16
 #sar 0: 1/1
-0,          0,          0,        1,     1024, 0x23e072c8
+0,          0,          0,        1,     1024, 0xc54d773d
-- 
2.39.0

_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".

  parent reply	other threads:[~2022-12-27 23:19 UTC|newest]

Thread overview: 75+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2022-11-05 15:26 [FFmpeg-devel] Rework color quantization in palette{gen,use} Clément Bœsch
2022-11-05 15:26 ` [FFmpeg-devel] [PATCH 01/15] Revert "avfilter/vf_palette(gen|use): support palettes with alpha" Clément Bœsch
2022-11-05 15:26 ` [FFmpeg-devel] [PATCH 02/15] avfilter: add palette utils Clément Bœsch
2022-11-05 15:26 ` [FFmpeg-devel] [PATCH 03/15] avfilter/palette{use, gen}: simplify a few alpha masks Clément Bœsch
2022-11-05 15:26 ` [FFmpeg-devel] [PATCH 04/15] avfilter/paletteuse: switch from u8[4] to u32 for color code Clément Bœsch
2022-11-05 15:26 ` [FFmpeg-devel] [PATCH 05/15] avfilter/paletteuse: name target color arg consistently in colormap functions Clément Bœsch
2022-11-05 15:26 ` [FFmpeg-devel] [PATCH 06/15] avfilter/paletteuse: remove unused alpha split dimension Clément Bœsch
2022-11-05 15:26 ` [FFmpeg-devel] [PATCH 07/15] avfilter/paletteuse: remove redundant alpha condition Clément Bœsch
2022-11-05 15:26 ` [FFmpeg-devel] [PATCH 08/15] avfilter/paletteuse: switch to a perceptual model Clément Bœsch
2022-11-05 15:26 ` [FFmpeg-devel] [PATCH 09/15] avfilter/palettegen: average color in linear space Clément Bœsch
2022-11-05 15:39   ` Paul B Mahol
2022-11-05 18:50     ` Clément Bœsch
2022-11-05 15:26 ` [FFmpeg-devel] [PATCH 10/15] avfilter/palettegen: move box variance computation in a dedicated function Clément Bœsch
2022-11-05 15:26 ` [FFmpeg-devel] [PATCH 11/15] avfilter/palettegen: comment on the unnormalized variance Clément Bœsch
2022-11-05 15:26 ` [FFmpeg-devel] [PATCH 12/15] avfilter/palettegen: base split decision on a perceptual model Clément Bœsch
2022-11-05 19:07   ` Andreas Rheinhardt
2022-11-08 21:09     ` Clément Bœsch
2022-12-27 23:20     ` Clément Bœsch
2022-11-05 15:26 ` [FFmpeg-devel] [PATCH 13/15] avfilter/palettegen: use variance per-axis instead of the range Clément Bœsch
2022-11-05 15:26 ` [FFmpeg-devel] [PATCH 14/15] avfilter/palettegen: rename longest to split_axis Clément Bœsch
2022-11-05 15:26 ` [FFmpeg-devel] [PATCH 15/15] avfilter/palette{use, gen}: update Copyright after recent changes Clément Bœsch
2022-11-05 15:44 ` [FFmpeg-devel] Rework color quantization in palette{gen,use} Paul B Mahol
2022-11-05 18:54   ` Clément Bœsch
2022-11-06 13:19     ` Ronald S. Bultje
2022-11-08 21:22       ` Clément Bœsch
2022-11-05 21:52 ` Soft Works
2022-11-06 17:09 ` Michael Niedermayer
2022-11-06 17:30   ` Michael Niedermayer
2022-11-08 21:14     ` Clément Bœsch
2022-12-31 12:11     ` Clément Bœsch
2023-01-02 21:57       ` Michael Niedermayer
2023-01-02 23:05         ` Clément Bœsch
2023-01-03 18:50           ` Michael Niedermayer
2022-11-06 19:46 ` Soft Works
2022-11-08 21:07   ` Clément Bœsch
2022-11-08 22:37     ` Soft Works
2022-12-27 23:31       ` Clément Bœsch
2022-12-27 23:17 ` [FFmpeg-devel] New iteration for the color quantization in palette{gen, use} Clément Bœsch
2022-12-27 23:17   ` [FFmpeg-devel] [PATCH v2 01/32] avfilter/palettegen: allow a minimum of 2 colors Clément Bœsch
2022-12-28 21:04     ` Tomas Härdin
2022-12-28 21:23       ` Clément Bœsch
2023-01-03 18:59         ` Tomas Härdin
2022-12-27 23:17   ` [FFmpeg-devel] [PATCH v2 02/32] avfilter/palette{gen, use}: revert support palettes with alpha Clément Bœsch
2023-01-03 19:11     ` Paul B Mahol
2022-12-27 23:17   ` [FFmpeg-devel] [PATCH v2 03/32] avfilter/palette{gen, use}: simplify a few alpha masks Clément Bœsch
2022-12-27 23:17   ` [FFmpeg-devel] [PATCH v2 04/32] avfilter/palette{gen, use}: add palette utils Clément Bœsch
2022-12-27 23:17   ` [FFmpeg-devel] [PATCH v2 05/32] avfilter/paletteuse: switch from u8[4] to u32 for color code Clément Bœsch
2022-12-27 23:17   ` [FFmpeg-devel] [PATCH v2 06/32] avfilter/paletteuse: name target color arg consistently in colormap functions Clément Bœsch
2022-12-27 23:17   ` [FFmpeg-devel] [PATCH v2 07/32] avfilter/paletteuse: remove unused alpha split dimension Clément Bœsch
2022-12-27 23:17   ` [FFmpeg-devel] [PATCH v2 08/32] avfilter/paletteuse: remove redundant alpha condition Clément Bœsch
2022-12-27 23:17   ` [FFmpeg-devel] [PATCH v2 09/32] avfilter/paletteuse: switch to a perceptual model Clément Bœsch
2022-12-27 23:17   ` [FFmpeg-devel] [PATCH v2 10/32] avfilter/palettegen: move box stats computation to a dedicated function Clément Bœsch
2022-12-27 23:17   ` Clément Bœsch [this message]
2022-12-27 23:17   ` [FFmpeg-devel] [PATCH v2 12/32] avfilter/palettegen: use box->major_axis without intermediate variable Clément Bœsch
2022-12-27 23:17   ` [FFmpeg-devel] [PATCH v2 13/32] avfilter/palettegen: always compute the box variance Clément Bœsch
2022-12-27 23:17   ` [FFmpeg-devel] [PATCH v2 14/32] avfilter/palettegen: rename variance to cut_score Clément Bœsch
2022-12-27 23:17   ` [FFmpeg-devel] [PATCH v2 15/32] avfilter/palettegen: change cut score from ∑e² to max e² Clément Bœsch
2022-12-27 23:17   ` [FFmpeg-devel] [PATCH v2 16/32] avfilter/palettegen: compute average color within compute_box_stats() Clément Bœsch
2022-12-27 23:17   ` [FFmpeg-devel] [PATCH v2 17/32] avfilter/palettegen: misc cosmetics Clément Bœsch
2022-12-27 23:18   ` [FFmpeg-devel] [PATCH v2 18/32] avfilter/palettegen: rename local variable box_weight to weight Clément Bœsch
2022-12-27 23:18   ` [FFmpeg-devel] [PATCH v2 19/32] avfilter/palettegen: switch to signed arithmetic Clément Bœsch
2022-12-27 23:18   ` [FFmpeg-devel] [PATCH v2 20/32] avfilter/palettegen: base box split decision on a perceptual model Clément Bœsch
2022-12-27 23:18   ` [FFmpeg-devel] [PATCH v2 21/32] avfilter/palettegen: add a warning about supporting only sRGB Clément Bœsch
2022-12-27 23:18   ` [FFmpeg-devel] [PATCH v2 22/32] avfilter/palettegen: make refs order deterministic Clément Bœsch
2022-12-27 23:18   ` [FFmpeg-devel] [PATCH v2 23/32] avfilter/palettegen: use libc qsort Clément Bœsch
2022-12-27 23:18   ` [FFmpeg-devel] [PATCH v2 24/32] avfilter/palette{gen, use}: update Copyright after recent changes Clément Bœsch
2022-12-27 23:18   ` [FFmpeg-devel] [PATCH v2 25/32] avfilter/palette: add lowbias32 hashing Clément Bœsch
2022-12-27 23:18   ` [FFmpeg-devel] [PATCH v2 26/32] avfilter/palettegen: use lowbias32 for color hashing Clément Bœsch
2022-12-27 23:18   ` [FFmpeg-devel] [PATCH v2 27/32] avfilter/paletteuse: " Clément Bœsch
2022-12-27 23:18   ` [FFmpeg-devel] [PATCH v2 28/32] avfilter/paletteuse: switch to recursive method Clément Bœsch
2022-12-27 23:18   ` [FFmpeg-devel] [PATCH v2 29/32] avfilter/paletteuse: remove alternative search methods Clément Bœsch
2022-12-27 23:18   ` [FFmpeg-devel] [PATCH v2 30/32] avfilter/paletteuse: remove mean error tool Clément Bœsch
2022-12-27 23:18   ` [FFmpeg-devel] [PATCH v2 31/32] avfilter/paletteuse: move r, g, b computation in a more local scope Clément Bœsch
2022-12-27 23:18   ` [FFmpeg-devel] [PATCH v2 32/32] avfilter/palette{gen, use}: misc for-loop cosmetics Clément Bœsch
2023-01-03 16:28   ` [FFmpeg-devel] New iteration for the color quantization in palette{gen, use} Clément Bœsch

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20221227231814.2520181-12-u@pkh.me \
    --to=u@pkh.me \
    --cc=ffmpeg-devel@ffmpeg.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

Git Inbox Mirror of the ffmpeg-devel mailing list - see https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

This inbox may be cloned and mirrored by anyone:

	git clone --mirror https://master.gitmailbox.com/ffmpegdev/0 ffmpegdev/git/0.git

	# If you have public-inbox 1.1+ installed, you may
	# initialize and index your mirror using the following commands:
	public-inbox-init -V2 ffmpegdev ffmpegdev/ https://master.gitmailbox.com/ffmpegdev \
		ffmpegdev@gitmailbox.com
	public-inbox-index ffmpegdev

Example config snippet for mirrors.


AGPL code for this site: git clone https://public-inbox.org/public-inbox.git