Inconsistent Results Across Different Operating Systems and Browsers with SsdMobilenetv1 Model

[IgnaValdiviaUTN]

Hi,

I'm experiencing inconsistent results when using the SsdMobilenetv1 model with face-api.js across different operating systems and browsers.

Comparativa face-api,js.pdf

Specifically, I'm running the following code to perform face recognition:

$(document).ready(function () {

        Promise.all([
            faceapi.nets.ssdMobilenetv1.loadFromUri("models/"),
            faceapi.nets.faceLandmark68Net.loadFromUri("models/"),
            faceapi.nets.faceRecognitionNet.loadFromUri("models/"),
            faceapi.nets.faceExpressionNet.loadFromUri("models/"),
            faceapi.nets.ageGenderNet.loadFromUri("models/"),
        ]).then(() => {
            console.log("Modelos cargados");
        });

        $("#enviar").click(async function () {
            
            const imagePath = "test/D6.png";

            
            const reconocimiento = {
                felicidad: 0,
                tristeza: 0,
                enojo: 0,
                temor: 0,
                neutralidad: 0,
                disgusto: 0,
                sorpresa: 0,
                edad: 0,
            };

            await procesarModeloJS(imagePath, reconocimiento);
        });
    });

    async function procesarModeloJS(imagePath, reconocimiento) {
        const img = new Image();
        img.src = imagePath;

        return new Promise((resolve) => {
            img.onload = async () => {
                try {
                    let predictions;

                    predictions = await faceapi
                        .detectSingleFace(img, new faceapi.SsdMobilenetv1Options())
                        .withFaceLandmarks()
                        .withFaceExpressions()
                        .withAgeAndGender();

                    if (predictions) {
                        Object.assign(reconocimiento, {
                            felicidad: predictions.expressions.happy * 100,
                            disgusto: predictions.expressions.disgusted * 100,
                            temor: predictions.expressions.fearful * 100,
                            enojo: predictions.expressions.angry * 100,
                            neutralidad: predictions.expressions.neutral * 100,
                            tristeza: predictions.expressions.sad * 100,
                            sorpresa: predictions.expressions.surprised * 100,
                            edad: predictions.age,
                        });

                    }
                } catch (error) {
                    console.error("Error en la predicción:", error);
                }

                resolve();
            };
        });
    }

The results vary significantly depending on the operating system and browser being used. For example, the accuracy of face detection, landmarks, expressions, age, and gender estimation differs between Windows and macOS, as well as between Chrome and Firefox.

Could you please provide some insights into why these discrepancies occur? Are there any known issues or limitations when using this model across different environments? Additionally, are there any recommended practices to ensure more consistent results?

Thank you for your help!