[FFmpeg-devel] [PATCH 09/11] tests/swscale: calculate theoretical expected SSIM

[Niklas Haas]

On Mon, 17 Mar 2025 11:43:55 +0100 Niklas Haas <ffmpeg at haasn.xyz> wrote:
> From: Niklas Haas <git at haasn.dev>
>
> We can calculate with some confidence the theoretical expected SSIM
> from an "ideal" conversion, by computing the reference SSIM level
> for an image dithered with uniformly distributed quatization noise.
>
> This gives us an additional safety net to check for regressions even in
> the absence of a reference to compare against.

It's worth pointing out that this does reveal some bugs in the current
implementation that were not covered by any pre existing tests.

> ---
>  libswscale/tests/swscale.c | 74 +++++++++++++++++++++++++++++++-------
>  1 file changed, 62 insertions(+), 12 deletions(-)
>
> diff --git a/libswscale/tests/swscale.c b/libswscale/tests/swscale.c
> index 47c58524f6..bce495db90 100644
> --- a/libswscale/tests/swscale.c
> +++ b/libswscale/tests/swscale.c
> @@ -99,6 +99,29 @@ static void exit_handler(int sig)
>      exit(sig);
>  }
>
> +/* Estimate luma variance assuming uniform dither noise distribution */
> +static float estimate_quantization_noise(enum AVPixelFormat fmt)
> +{
> +    const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(fmt);
> +    float variance = 1.0 / 12;
> +    if (desc->comp[0].depth < 8) {
> +        /* Extra headroom for very low bit depth output */
> +        variance *= (8 - desc->comp[0].depth);
> +    }
> +
> +    if (desc->flags & AV_PIX_FMT_FLAG_FLOAT) {
> +        return 0.0;
> +    } else if (desc->flags & AV_PIX_FMT_FLAG_RGB) {
> +        const float r = 0.299 / (1 << desc->comp[0].depth);
> +        const float g = 0.587 / (1 << desc->comp[1].depth);
> +        const float b = 0.114 / (1 << desc->comp[2].depth);
> +        return (r * r + g * g + b * b) * variance;
> +    } else {
> +        const float y = 1.0 / (1 << desc->comp[0].depth);
> +        return y * y * variance;
> +    }
> +}
> +
>  static int fmt_comps(enum AVPixelFormat fmt)
>  {
>      const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(fmt);
> @@ -156,6 +179,18 @@ static void get_ssim(float ssim[4], const AVFrame *out, const AVFrame *ref, int
>      }
>  }
>
> +static float get_loss(const float ssim[4])
> +{
> +    const float weights[3] = { 0.8, 0.1, 0.1 }; /* tuned for Y'CrCr */
> +
> +    float sum = 0;
> +    for (int i = 0; i < 3; i++)
> +        sum += weights[i] * ssim[i];
> +    sum *= ssim[3]; /* ensure alpha errors get caught */
> +
> +    return 1.0 - sum;
> +}
> +
>  static int scale_legacy(AVFrame *dst, const AVFrame *src, struct mode mode,
>                          struct options opts)
>  {
> @@ -198,6 +233,18 @@ static int run_test(enum AVPixelFormat src_fmt, enum AVPixelFormat dst_fmt,
>      int64_t time, time_ref = 0;
>      int ret = -1;
>
> +    /* Estimate the expected amount of loss from bit depth reduction */
> +    const float c1 = 0.01 * 0.01; /* stabilization constant */
> +    const float ref_var = 1.0 / 12.0; /* uniformly distributed signal */
> +    const float src_var = estimate_quantization_noise(src_fmt);
> +    const float dst_var = estimate_quantization_noise(dst_fmt);
> +    const float out_var = estimate_quantization_noise(ref->format);
> +    const float total_var = src_var + dst_var + out_var;
> +    const float ssim_luma = (2 * ref_var + c1) / (2 * ref_var + total_var + c1);
> +    const float ssim_expected[4] = { ssim_luma, 1, 1, 1 }; /* for simplicity */
> +    const float expected_loss = get_loss(ssim_expected);
> +    float loss;
> +
>      src = av_frame_alloc();
>      dst = av_frame_alloc();
>      out = av_frame_alloc();
> @@ -251,6 +298,15 @@ static int run_test(enum AVPixelFormat src_fmt, enum AVPixelFormat dst_fmt,
>             mode.flags, mode.dither,
>             ssim[0], ssim[1], ssim[2], ssim[3]);
>
> +    loss = get_loss(ssim);
> +    if (loss - expected_loss > 1e-4 && dst_w >= ref->width && dst_h >= ref->height) {
> +        int bad = loss - expected_loss > 1e-2;
> +        printf("\033[1;31m  loss %g is %s by %g, expected loss %g\033[0m\n",
> +               loss, bad ? "WORSE" : "worse", loss - expected_loss, expected_loss);
> +        if (bad)
> +            goto error;
> +    }
> +
>      if (!ssim_ref && sws_isSupportedInput(src->format) && sws_isSupportedOutput(dst->format)) {
>          /* Compare against the legacy swscale API as a reference */
>          time_ref = av_gettime_relative();
> @@ -269,18 +325,12 @@ static int run_test(enum AVPixelFormat src_fmt, enum AVPixelFormat dst_fmt,
>      }
>
>      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++) {
> -            sum     += weights[i] * ssim[i];
> -            sum_ref += weights[i] * ssim_ref[i];
> -        }
> -
> -        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,
> +        const float loss_ref = get_loss(ssim_ref);
> +        if (loss - loss_ref > 1e-4) {
> +            int bad = loss - loss_ref > 1e-2;
> +            printf("\033[1;31m  loss %g is %s by %g, ref loss %g, "
> +                   "SSIM {Y=%f U=%f V=%f A=%f}\033[0m\n",
> +                   loss, bad ? "WORSE" : "worse", loss - loss_ref, loss_ref,
>                     ssim_ref[0], ssim_ref[1], ssim_ref[2], ssim_ref[3]);
>              if (bad)
>                  goto error;
> --
> 2.48.1
>