About Confidence Value Range

[burcinsahin]

Hi georg-jung,

I am trying face detection and alignment basically for one image. Everything works as it expected to be. It detects face and aligns it perfectly. I really appreciate the work you did.

But the confidence value is higher than 1 (6.2194686). So I am confused here. I thought confidence value should be some value between '0.0' and '1.0'. So am I wrong here? Is it normal? I could not find any information about confidence value in readme part. So I created a new issue to get an answer.

I added a screenshot of quick watch as attachment.

My test method is like below:

Thanks in advance.
Best Regards.

[00f8t]

any solution?

[georg-jung]

Sorry for the delay. I also think confidence should be in the range [0 .. 1]. Which values do you get with the exact image from the readme? Did you experience this with multiple images or only one specific picture? Can you share the picture so that I can reproduce this behavior?

[00f8t]

Sorry for the delay. I also think confidence should be in the range [0 .. 1]. Which values do you get with the exact image from the readme? Did you experience this with multiple images or only one specific picture? Can you share the picture so that I can reproduce this behavior?

i tested it, turns out that it depends on input image size, this value becomes as many times larger as the size of the original image is larger than the one expected by the model

[georg-jung]

Interesting. FaceAiSharp is meant to resize to an appropriate size before feeding the image to the model. Can you add more details about your test setup and/or the image you tested with so that I can reproduce this?

[burcinsahin]

Hi, I experienced the issue with 2 images. I can not share them because they contain private data. Sorry for that. I will try to create the same issue with some unimportant images and send to you. Thanks.

[burcinsahin]

Hi, I found a public sample image and it gives a confidence value more than 1. Detection and aligning work as expected. You can try this one. Thanks

Source: https://x.com/vekilince/status/1456969787897229320

[00f8t]

Hi, I think I found the source of an issue in the post-processing logic.

In ScrfdDetector.cs, lines 212–248, you're scaling the bounding box coordinates like this:

var det = preDet[keep];
det *= scale;

NDArray? kpss = null;
if (kpssLst.Count > 0)
{
    kpss = np.vstack(kpssLst.ToArray());
    kpss = kpss[order];
    kpss = kpss[keep];
    kpss *= scale;
}

Later, when returning the detection results, you're extracting the confidence score from the same scaled NDArray:

static FaceDetectorResult ToReturnType(NDArray input)
{
    var x1 = input.GetSingle(0);
    var y1 = input.GetSingle(1);
    var x2 = input.GetSingle(2);
    var y2 = input.GetSingle(3);
    return new(new RectangleF(x1, y1, x2 - x1, y2 - y1), null, input.GetSingle(4));
}

Since det *= scale affects the entire NDArray, this also unintentionally scales the confidence value. This caused issues for me when processing images of different sizes. As a temporary workaround, I divided the confidence by the scale factor afterward, and it fixed the problem:

static FaceDetectorResult ToReturnTypeWithLandmarks(NDArray input, NDArray kps, float scale)
{
    var (box, _, conf) = ToReturnType(input);
    var lmrks = new List<PointF>(5);
    lmrks.AddRange(kps.GetNDArrays(0).Select(x => new PointF(x.GetSingle(0), x.GetSingle(1))));
    return new(box, lmrks, conf / scale);
}

Maybe it would be better to scale only the bounding box and landmark coordinates, not the entire detection array?

Unrelated note: I'm also using byte array instead of SixLabors.ImageSharp.Image, thinking it might offer better performance and versatility for my use case.

[martindevans]

I just cloned the repo, intending to PR a fix, but I can't reproduce it. It looks like this may have been fixed internally, but hasn't been released to nuget yet.

If that's the case, could I request an update of the nuget package? Thanks ❤️

[georg-jung]

Which version did you try? Note that there is a more recent pre-release version. Might be about time to make it a stable release. I think I'll do so in the coming days.

[martindevans]

Aha I was trying with 0.5.23 which is the current stable version. I just updated to 0.6.25-alpha and the confidence issue seems to be fixed. Thanks for help.