virtualenv/FFmpeg
1
0
Fork 0
mirror of https://github.com/FFmpeg/FFmpeg.git synced 2025-08-10 06:10:52 +02:00
Code Issues Releases Activity

tests/swscale: switch from MSE to SSIM

Browse Source 
And bias it towards Y. This is much better at ignoring errors due to differing
dither patterns, and rewards algorithms that lower luma noise at the cost of
higher chroma noise.

The (0.8, 0.1, 0.1) weights for YCbCr are taken from the paper:
  "Understanding SSIM" by Jim Nilsson and Tomas Akenine-Möller
  (https://arxiv.org/abs/2006.13846)
This commit is contained in:
Niklas Haas
2025-02-25 15:47:43 +01:00
parent 1707e81073
commit 6c12b1535a
1 changed files with 65 additions and 43 deletions
Download Patch File Download Diff File Expand all files Collapse all files

106
libswscale/tests/swscale.c
View File

@@ -80,38 +80,51 @@ static int fmt_comps(enum AVPixelFormat fmt)
return comps; return comps;
} }
static void get_mse(int mse[4], const AVFrame *a, const AVFrame *b, int comps) static void get_ssim(float ssim[4], const AVFrame *out, const AVFrame *ref, int comps)
{ {
av_assert1(a->format == AV_PIX_FMT_YUVA444P); av_assert1(out->format == AV_PIX_FMT_YUVA444P);
av_assert1(b->format == a->format); av_assert1(ref->format == out->format);
av_assert1(b->width == a->width && b->height == a->height); av_assert1(ref->width == out->width && ref->height == out->height);
for (int p = 0; p < 4; p++) { for (int p = 0; p < 4; p++) {
const int stride_a = out->linesize[p];
const int stride_b = ref->linesize[p];
const int w = out->width;
const int h = out->height;
const int is_chroma = p == 1 || p == 2; const int is_chroma = p == 1 || p == 2;
const int stride_a = a->linesize[p]; const uint8_t def = is_chroma ? 128 : 0xFF;
const int stride_b = b->linesize[p]; const int has_ref = comps & (1 << p);
const int w = a->width; double sum = 0;
const int h = a->height; int count = 0;
uint64_t sum = 0;
if (comps & (1 << p)) { /* 4x4 SSIM */
for (int y = 0; y < h; y++) { for (int y = 0; y < (h & ~3); y += 4) {
for (int x = 0; x < w; x++) { for (int x = 0; x < (w & ~3); x += 4) {
int d = a->data[p][y * stride_a + x] - b->data[p][y * stride_b + x]; const float c1 = .01 * .01 * 255 * 255 * 64;
sum += d * d; const float c2 = .03 * .03 * 255 * 255 * 64 * 63;
} int s1 = 0, s2 = 0, ss = 0, s12 = 0, var, covar;
}
} else { for (int yy = 0; yy < 4; yy++) {
const int ref = is_chroma ? 128 : 0xFF; for (int xx = 0; xx < 4; xx++) {
for (int y = 0; y < h; y++) { int a = out->data[p][(y + yy) * stride_a + x + xx];
for (int x = 0; x < w; x++) { int b = has_ref ? ref->data[p][(y + yy) * stride_b + x + xx] : def;
int d = a->data[p][y * stride_a + x] - ref; s1 += a;
sum += d * d; s2 += b;
} ss += a * a + b * b;
s12 += a * b;
} }
} }
mse[p] = sum / (w * h); var = ss * 64 - s1 * s1 - s2 * s2;
covar = s12 * 64 - s1 * s2;
sum += (2 * s1 * s2 + c1) * (2 * covar + c2) /
((s1 * s1 + s2 * s2 + c1) * (var + c2));
count++;
}
}
ssim[p] = count ? sum / count : 0.0;
} }
} }
@@ -149,12 +162,13 @@ error:
/* Runs a series of ref -> src -> dst -> out, and compares out vs ref */ /* Runs a series of ref -> src -> dst -> out, and compares out vs ref */
static int run_test(enum AVPixelFormat src_fmt, enum AVPixelFormat dst_fmt, static int run_test(enum AVPixelFormat src_fmt, enum AVPixelFormat dst_fmt,
int dst_w, int dst_h, struct mode mode, struct options opts, int dst_w, int dst_h, struct mode mode, struct options opts,
const AVFrame *ref, const int mse_ref[4]) const AVFrame *ref, const float ssim_ref[4])
{ {
AVFrame *src = NULL, *dst = NULL, *out = NULL; AVFrame *src = NULL, *dst = NULL, *out = NULL;
int mse[4], mse_sws[4], ret = -1; float ssim[4], ssim_sws[4];
const int comps = fmt_comps(src_fmt) & fmt_comps(dst_fmt); const int comps = fmt_comps(src_fmt) & fmt_comps(dst_fmt);
int64_t time, time_ref = 0; int64_t time, time_ref = 0;
int ret = -1;
src = av_frame_alloc(); src = av_frame_alloc();
dst = av_frame_alloc(); dst = av_frame_alloc();
@@ -201,15 +215,15 @@ static int run_test(enum AVPixelFormat src_fmt, enum AVPixelFormat dst_fmt,
goto error; goto error;
} }
get_mse(mse, out, ref, comps); get_ssim(ssim, out, ref, comps);
printf("%s %dx%d -> %s %3dx%3d, flags=0x%x dither=%u, " printf("%s %dx%d -> %s %3dx%3d, flags=0x%x dither=%u, "
"MSE={%5d %5d %5d %5d}\n", "SSIM {Y=%f U=%f V=%f A=%f}\n",
av_get_pix_fmt_name(src->format), src->width, src->height, av_get_pix_fmt_name(src->format), src->width, src->height,
av_get_pix_fmt_name(dst->format), dst->width, dst->height, av_get_pix_fmt_name(dst->format), dst->width, dst->height,
mode.flags, mode.dither, mode.flags, mode.dither,
mse[0], mse[1], mse[2], mse[3]); ssim[0], ssim[1], ssim[2], ssim[3]);
if (!mse_ref) { if (!ssim_ref) {
/* Compare against the legacy swscale API as a reference */ /* Compare against the legacy swscale API as a reference */
time_ref = av_gettime_relative(); time_ref = av_gettime_relative();
if (scale_legacy(dst, src, mode, opts) < 0) { if (scale_legacy(dst, src, mode, opts) < 0) {
@@ -222,19 +236,26 @@ static int run_test(enum AVPixelFormat src_fmt, enum AVPixelFormat dst_fmt,
if (sws_scale_frame(sws[2], out, dst) < 0) if (sws_scale_frame(sws[2], out, dst) < 0)
goto error; goto error;
get_mse(mse_sws, out, ref, comps); get_ssim(ssim_sws, out, ref, comps);
mse_ref = mse_sws; ssim_ref = ssim_sws;
} }
if (ssim_ref) {
const float weights[4] = { 0.8, 0.1, 0.1, 1.0 }; /* tuned for Y'CrCr */
float err, sum = 0, sum_ref = 0;
for (int i = 0; i < 4; i++) { for (int i = 0; i < 4; i++) {
if (mse[i] > mse_ref[i]) { sum += weights[i] * ssim[i];
int bad = mse[i] > mse_ref[i] * 1.02 + 1; sum_ref += weights[i] * ssim_ref[i];
printf("\033[1;31m %s, ref MSE={%5d %5d %5d %5d}\033[0m\n", }
bad ? "WORSE" : "worse",
mse_ref[0], mse_ref[1], mse_ref[2], mse_ref[3]); err = sum_ref / sum - 1.0; /* relative error */
if (err > 1e-4 /* 0.01% headroom for dither noise etc */) {
int bad = err > 1e-2; /* 1% */
printf("\033[1;31m %s by %f%%, ref SSIM {Y=%f U=%f V=%f A=%f}\033[0m\n",
bad ? "WORSE" : "worse", 100.0 * err,
ssim_ref[0], ssim_ref[1], ssim_ref[2], ssim_ref[3]);
if (bad) if (bad)
goto error; goto error;
break;
} }
} }
@@ -334,15 +355,16 @@ static int run_file_tests(const AVFrame *ref, FILE *fp, struct options opts)
char src_fmt_str[21], dst_fmt_str[21]; char src_fmt_str[21], dst_fmt_str[21];
enum AVPixelFormat src_fmt; enum AVPixelFormat src_fmt;
enum AVPixelFormat dst_fmt; enum AVPixelFormat dst_fmt;
int sw, sh, dw, dh, mse[4]; int sw, sh, dw, dh;
float ssim[4];
struct mode mode; struct mode mode;
ret = sscanf(buf, ret = sscanf(buf,
" %20s %dx%d -> %20s %dx%d, flags=0x%x dither=%u, " " %20s %dx%d -> %20s %dx%d, flags=0x%x dither=%u, "
"MSE={%d %d %d %d}\n", "SSIM {Y=%f U=%f V=%f A=%f}\n",
src_fmt_str, &sw, &sh, dst_fmt_str, &dw, &dh, src_fmt_str, &sw, &sh, dst_fmt_str, &dw, &dh,
&mode.flags, &mode.dither, &mode.flags, &mode.dither,
&mse[0], &mse[1], &mse[2], &mse[3]); &ssim[0], &ssim[1], &ssim[2], &ssim[3]);
if (ret != 12) { if (ret != 12) {
printf("%s", buf); printf("%s", buf);
continue; continue;
@@ -361,7 +383,7 @@ static int run_file_tests(const AVFrame *ref, FILE *fp, struct options opts)
opts.dst_fmt != AV_PIX_FMT_NONE && dst_fmt != opts.dst_fmt) opts.dst_fmt != AV_PIX_FMT_NONE && dst_fmt != opts.dst_fmt)
continue; continue;
if (run_test(src_fmt, dst_fmt, dw, dh, mode, opts, ref, mse) < 0) if (run_test(src_fmt, dst_fmt, dw, dh, mode, opts, ref, ssim) < 0)
return -1; return -1;
} }