Better way to compare two faces?

[pvictorlv]

Hey, i need to compare two faces (without storing them, both in base64) and display the similarity level of them. What's the better way to do it? And how i can increase the precision level? Here is my code now:

    public async Task<double> CompareFaces(string sourceImageBase, string targetImageBase)
        {
            string directory = "C:\\models";

            using (FaceRecognition fr = FaceRecognition.Create(directory))
            {
                var bitmapData = GetImageBytes(FixBase64ForImage(sourceImageBase));
                var targetBitmapData = GetImageBytes(FixBase64ForImage(targetImageBase));

                await using (MemoryStream streamTargetBitmap = new MemoryStream(targetBitmapData))
                {
                    using (Bitmap bitTargetImage = new Bitmap(streamTargetBitmap))
                    {
                        await using (MemoryStream streamSourceBitmap = new MemoryStream(bitmapData))
                        {
                            using (Bitmap bitSourceImage = new Bitmap(streamSourceBitmap))
                            {
                                using (var imageA = FaceRecognition.LoadImage(bitSourceImage))
                                {
                                    using (var imageB = FaceRecognition.LoadImage(bitTargetImage))
                                    {
                                        IEnumerable<Location> locationsA = fr.FaceLocations(imageA);
                                        IEnumerable<Location> locationsB = fr.FaceLocations(imageB);

                                        var encodingA = fr.FaceEncodings(imageA, locationsA);
                                        IEnumerable<FaceEncoding> encodingB = fr.FaceEncodings(imageB, locationsB);

                                        return FaceRecognition.FaceDistance(encodingA.First(), encodingB.First());
                                    }
                                }
                            }
                        }
                    }
                }
            }
        }

Thanks!

[takuya-takeuchi]

To increase accuracy

• Use large image
• Use front face image

What's the better way to do it?

And you need not to re-create FaceRecognition object. You can share it on 1-thread. If you want to use multithread, you must create N FaceRecognition object for N threads.

[pvictorlv]

I've tried with 100x100 grayscale face cut out image and with normal images in 520x360 with same results, and it's bad. it can't identify the same person... btw, how the distance is calculated? it's from 0 to 100?

[pvictorlv]

@takuya-takeuchi How can i contact you?

[takuya-takeuchi]

100x100 grayscale

You mean that you compare with encoding from grayscale and encoding from rgb image, right? It could be no matter but I'm not sure.

btw, how the distance is calculated? it's from 0 to 100?

dittance is from 0 to 1.0.

How can i contact you?

No. I am very very busy and I can not contribute to my projects.

[pvictorlv]

Btw, there's any way to improve the recognition speed?

[takuya-takeuchi]

use gpu or intel MKL