Assets/Scripts/BlazeUtils.cs

using System;
using System.Globalization;
using Unity.Mathematics;
using Unity.Sentis;
using UnityEngine;

public static class BlazeUtils
{
    // matrix utility
    public static float2x3 mul(float2x3 a, float2x3 b)
    {
        return new float2x3(
            a[0][0] * b[0][0] + a[1][0] * b[0][1],
            a[0][0] * b[1][0] + a[1][0] * b[1][1],
            a[0][0] * b[2][0] + a[1][0] * b[2][1] + a[2][0],
            a[0][1] * b[0][0] + a[1][1] * b[0][1],
            a[0][1] * b[1][0] + a[1][1] * b[1][1],
            a[0][1] * b[2][0] + a[1][1] * b[2][1] + a[2][1]
        );
    }

    public static float2 mul(float2x3 a, float2 b)
    {
        return new float2(
            a[0][0] * b.x + a[1][0] * b.y + a[2][0],
            a[0][1] * b.x + a[1][1] * b.y + a[2][1]
        );
    }

    public static float2x3 RotationMatrix(float theta)
    {
        var sinTheta = math.sin(theta);
        var cosTheta = math.cos(theta);
        return new float2x3(
            cosTheta, -sinTheta, 0,
            sinTheta, cosTheta, 0
        );
    }

    public static float2x3 TranslationMatrix(float2 delta)
    {
        return new float2x3(
            1, 0, delta.x,
            0, 1, delta.y
        );
    }

    public static float2x3 ScaleMatrix(float2 scale)
    {
        return new float2x3(
            scale.x, 0, 0,
            0, scale.y, 0
        );
    }

    // model filtering utility
    static FunctionalTensor ScoreFiltering(FunctionalTensor rawScores, float scoreThreshold)
    {
        return Functional.Sigmoid(Functional.Clamp(rawScores, -scoreThreshold, scoreThreshold));
    }

    public static (FunctionalTensor, FunctionalTensor, FunctionalTensor) ArgMaxFiltering(FunctionalTensor rawBoxes, FunctionalTensor rawScores)
    {
        var detectionScores = ScoreFiltering(rawScores, 100f); // (1, 2254, 1)
        var bestScoreIndex = Functional.ArgMax(rawScores, 1).Squeeze();

        var selectedBoxes = Functional.IndexSelect(rawBoxes, 1, bestScoreIndex).Unsqueeze(0); // (1, 1, 16)
        var selectedScores = Functional.IndexSelect(detectionScores, 1, bestScoreIndex).Unsqueeze(0); // (1, 1, 1)

        return (bestScoreIndex, selectedScores, selectedBoxes);
    }

    // image transform utility
    static ComputeShader s_ImageTransformShader = Resources.Load<ComputeShader>("ComputeShaders/ImageTransform");
    static int s_ImageSample = s_ImageTransformShader.FindKernel("ImageSample");
    static int s_Optr = Shader.PropertyToID("Optr");
    static int s_X_tex2D = Shader.PropertyToID("X_tex2D");
    static int s_O_height = Shader.PropertyToID("O_height");
    static int s_O_width = Shader.PropertyToID("O_width");
    static int s_O_channels = Shader.PropertyToID("O_channels");
    static int s_X_height = Shader.PropertyToID("X_height");
    static int s_X_width = Shader.PropertyToID("X_width");
    static int s_affineMatrix = Shader.PropertyToID("affineMatrix");

    static int IDivC(int v, int div)
    {
        return (v + div - 1) / div;
    }

    public static void SampleImageAffine(Texture srcTexture, Tensor<float> dstTensor, float2x3 M)
    {
        var tensorData = ComputeTensorData.Pin(dstTensor, false);

        s_ImageTransformShader.SetTexture(s_ImageSample, s_X_tex2D, srcTexture);
        s_ImageTransformShader.SetBuffer(s_ImageSample, s_Optr, tensorData.buffer);

        s_ImageTransformShader.SetInt(s_O_height, dstTensor.shape[1]);
        s_ImageTransformShader.SetInt(s_O_width, dstTensor.shape[2]);
        s_ImageTransformShader.SetInt(s_O_channels, dstTensor.shape[3]);
        s_ImageTransformShader.SetInt(s_X_height, srcTexture.height);
        s_ImageTransformShader.SetInt(s_X_width, srcTexture.width);

        s_ImageTransformShader.SetMatrix(s_affineMatrix, new Matrix4x4(new Vector4(M[0][0], M[0][1]), new Vector4(M[1][0], M[1][1]), new Vector4(M[2][0], M[2][1]), Vector4.zero));

        s_ImageTransformShader.Dispatch(s_ImageSample, IDivC(dstTensor.shape[1], 8), IDivC(dstTensor.shape[1], 8), 1);
    }

    public static float[,] LoadAnchors(string csv, int numAnchors)
    {
        var anchors = new float[numAnchors, 4];
        var anchorLines = csv.Split('\n');

        for (var i = 0; i < numAnchors; i++)
        {
            var anchorValues = anchorLines[i].Split(',');
            for (var j = 0; j < 4; j++)
            {
                anchors[i, j] = float.Parse(anchorValues[j], CultureInfo.InvariantCulture);
            }
        }

        return anchors;
    }
}