Need tool to score output on psychovisual equivalency versus other renderers.

[ILOVEPIE]

This should use a still image perceptual comparison algorithm to compare output. Something similar to PSNR.

[abhishekmg]

@ILOVEPIE , please do check out this function in this JS playground https://playcode.io/1752106

if the RGB values match then we get a value as Infinity

Please do let me know if this is what you are looking for, if not any inputs on what exactly needed is much appreciated !

If this is what you are looking for I would be much interested in raising a PR !

[ILOVEPIE]

Well, you want to do PSNR or SSIM on XYZ colors, not RGB, also your code isn't taking the Mean Squared Error correctly.

This is more what I'm talking about:

function multiplyMatrixByVector(matrix, vector) {
    let result = [0, 0, 0];
    for (let i = 0; i < 3; i++) {
        for (let j = 0; j < 3; j++) {
            result[i] += matrix[i * 3 + j] * vector[j];
        }
    }
    
    return result;
}

function sRGBtoRGB(sR,sG,sB){
	const color = [sR,sG,sB];
  for(let i = 0; i < 3; i++){
  	if(color[i] <= 0.04045){
        color[i] /= 12.92;
    }else{
        color[i] = Math.pow((color[i]+0.055)/1.055,2.4);
    }
  }
  return color;
}
const DP3toLinearDP3 = sRGBtoRGB;

function sRGBtoXYZ(sR,sG,sB){
	const RGBtoXYZMatrix = [0.4360413,0.3851129,0.1430458,0.2224845,0.7169051,0.0606104,0.0139202,0.0970672,0.7139126];
	const RGB = sRGBtoRGB(sR,sG,sB);
  return multiplyMatrixByVector(RGBtoXYZMatrix,RGB);
}

function DP3toXYZ(wR,wG,wB){
	const LinearDP3toXYZMatrix = [0.5151187,0.2919778,0.1571035,0.2411892,0.6922441,0.0665668,-0.0010505,0.0418791,0.7840713];
	const LinearDP3 = DP3toLinearDP3(wR,wG,wB)
  return multiplyMatrixByVector(LinearDP3toXYZMatrix,LinearDP3)
}

function imageDataToXYZ(d){
	const data = d.data;
  const colorSpace = d.colorSpace;
  const length = d.width*d.height*3;
  const output = new Array(length);
  for(let i = 0; i < length; i+=3){
		let XYZ;
    switch(colorSpace){
    	default:
				console.warn("Unrecognized color space '"+colorSpace+"' assuming srgb, colors may be wrong.");
			case "srgb":
      	XYZ = sRGBtoXYZ(data[0]/255,data[1]/255,data[2]/255);
        break;
      case "display-p3":
				XYZ = DP3toXYZ(data[0]/255,data[1]/255,data[2]/255);
        break;
		}
    output[i] = XYZ[0];
    output[i+1] = XYZ[1];
    output[i+2] = XYZ[2];
	}
  return output;
}

function diffArray(a,b){
	const maxLength = Math.max(a.length, b.length);
  const output = new Array(maxLength);
  for(let i = 0; i < maxLength; i++){
		output[i] = (a[i]||0)-(b[i]||0);
  }
  return output;
}

function selectArrayFromArray(a,n,x){
	const offset = typeof(x) === "number" ? x : 0;
	const length = Math.trunc(a.length/n);
  const output = new Array(length);
  for(let i = 0; i < length; i++){
		output[i] = a[(i*n)+offset];
  }
  return output;
}

function psnr(image1,image2) {
		const image1xyz = imageDataToXYZ(image1);
    const image2xyz = imageDataToXYZ(image2);
    
    let mse;
   	{
    	const xyzDiffSquared = diffArray(image1xyz,image2xyz).map((value) => Math.pow(value,2));
      const mseComponentLength = Math.trunc(xyzDiffSquared.length/3);
      // Calculate the mean squared error for each channel
      const mseX = selectArrayFromArray(xyzDiffSquared,3).reduce((sum,value) => sum+value,0)/mseComponentLength;
      const mseY = selectArrayFromArray(xyzDiffSquared,3,1).reduce((sum,value) => sum+value,0)/mseComponentLength;
      const mseZ = selectArrayFromArray(xyzDiffSquared,3,2).reduce((sum,value) => sum+value,0)/mseComponentLength;
      // Calculate the mean squared error across all channels
      mse = (mseX + mseY + mseZ) / 3;
    }

    // Maximum possible pixel value (we are using floating point values)
    const maxPixelValue = 1.0;

		// Return PSNR value.
    return 20 * Math.log10(maxPixelValue / Math.sqrt(mse));
}

[abhishekmg]

Right Thank you for the input ! Things look a bit complex to me 😆 still need to look into this to understand it better

But it looks like this is the function you are looking for, so this issue can be closed ?

[ILOVEPIE]

But it looks like this is the function you are looking for, so this issue can be closed ?

Not exactly, I was looking for someone to integrate this into the unit test suite.

[abhishekmg]

Like having the function that you mentioned here #66 (comment) as a helper function in src/__tests__/test-utils folder ?

[ILOVEPIE]

yeah but we need to be able to use libass to get a comparison.

[ILOVEPIE]

I've coded up a working implementation that does PSNR and SSIM, but we need to convert it from a webpage to a node tool that can be used to compare with output from libass.