Dancing Glow

Dancing Glow

Oct 7, 2019

This week I tried real-time data collection and multiple, different types of inputs and output with the new Neural Network feature of ml5.js. Dancing Glow, as a result, is a visual project in which lines following the body will change to different colors according to body positions as trained. Here's the LINK to the p5 web editor.

Train

Data Collection

I set up and used two models in this project - one is 'regression' model that trains the program to predict the color of any given poses; another one is 'classification' model that trains the program to distinguish normal poses to emit pose, when posing which the glowing colors will emit to the surroundings. To collect data for both models, the user need to add data with either a color array as target output, or a label of "Emit" or "Not Emit". Moreover, to collect data more efficiently, while data is collected to color regression model, the same pose information is also passed into the classification model as “Not Emit” label.

//  Add a training example
function addExample_1() {
    // Add to brain_1 as color input
    // Add to brain_2 as Not Emit category
    if (poses.length > 0) {
        let poseInput = get_body_data(poses);
        let target = bodyColor;
        // Add data for regression training
        // Input 26, Output 3
        brain_1.addData(poseInput, [target[0], target[1], target[2]]);
        // Add data for categorization training
        // Input 26, Output 1
        brain_2.addData(poseInput, ["Not Emit"]);
    }
}

function addExample_2() {
    // Add to brain_2 as Emit category
    if (poses.length > 0) {
        let poseInput = get_body_data(poses);
        brain_2.addData(poseInput, ["Emit"]);
    }
}

Mouth Switch


Face Switch

Another interesting thing, which you might have noticed, is that I open and close my mouth as the switch of data collection. We all know about the 3..2..1.. countdown, which is decent and boring. Here instead I use face-api.js (a built-in library of ml5.js) to track the points of my mouth and calculate the distances of several signature points to get to know whether I'm opening my mouth or not, thus if to collect the pose data.

poseNet = ml5.poseNet(video, modelReady);
// This sets up an event that fills the global variable "poses"
// with an array every time new poses are detected
poseNet.on("pose", function (results) {
    poses = results;
});

// face-api.js
// No need to create faceBrain here
const faceOptions = {
    withLandmarks: true,
    withExpressions: false,
    withDescriptors: false
};
faceStatusLabel = 2;
faceapi = ml5.faceApi(video, faceOptions, faceReady);

Here poseNet is for tracking poses as inputs for machine learning, and faceAPI is solely for mouth opening and closing. The problem is that face-api.js is really heavy and slows down the frame rate significantly once it's running. To save computing resources, instead of letting it recursively call faceapi.detect(gotFaces), I move it to draw() and will only be executed when the user choose a color or emit class to collect data.

Classification

However, the classification model is not working, getting weird predictions in which one label's confidence is always 1 and another one's is undefined.

Array(2)
0: {label: "Emit", confidence: 1}
1: {label: "Not Emit", confidence: undefined}
tensor: t {kept: false, isDisposedInternal: false, shape: Array(2), dtype: "float32", size: 1, …}
length: 2
__proto__: Array(0)

Interface

The Interface of Dancing Glow

The interface for user to collect data and train the model is another focus of the project. It's designed for users to easily select data target (for color regression or emit classification), adjust color, add more classes of data (user can add up to 8 colors), and get informed of the status behind the scene.

The Interface of Dancing Glow - Selections

At the top of the interface are add color class button for adding more color classes, add data for emit class button, add data for color class button, and train! button. Since the project involves multiple library and the training of more than one model, I added a section of status feedback for users to know about the progress more easily. Below the video are status reports of the loading of libraries and training of the model - grey means not start yet, yellow means in-progress, and green means finished.

Since the person collecting the data will pose and be away from the screen during the data collection, text or icon could be hard for them to read or distinguish. Thus a green tint would be applied to the whole video area once the model detected the mouth opening and data being collected.

Save and Upload

(Updated in Oct. 14th, 2019) Then I added the latest save/load data and model features of ml5.neuralNetwork. It would then be much easier for people to reuse collected pose data to try different training options or to use trained model on different devices.

Download Buttons

The ↓data button can be activated only when proper data are collected, and 2 files will be downloaded afterwards:

>>> From Brain 1
poses_data_for_color.json
>>> From Brain 2
poses_data_for_emit.json

The ↓model button can be activated only after the model is trained, and in this project, six files will be downloaded, 3 for each “brain”.

model_meta.json
model.json
model.weights.bin

Then I want to load the downloaded data to another window with a untrained model. Since I want the data also be reflected to the interface explicitly instead of just being added silently, after calling .loadData() function, I also have the program run through the original .json file and add them to the displayed classes.

Not working

What I want is when the user drop a .json file on the canvas, the program will read and load the file, load them into brain 1 or 2, and reflect on the interface. As the canvas.drop(gotFile) examples in p5 documentation, the code goes as:

function gotFile(file) {
    if (file.subtype === "json") {
        const dataFile = loadJSON(file.data);
        if (file.name == "poses_data_for_color.json") {
            brain_1.loadData(file.data, dataLoaded); // Not sure if is working

            console.log(dataFile); // Console print an object
            console.log(dataFile['data']); // Console print undefined
        } else if (file.name == "poses_data_for_emit.json") {
            brain_2.loadData(file.data, dataLoaded);
        } else {
            console.log('File Name Error');
        }
    } else {
        console.log('File Type Error');
    }
}

Here's what I got when console print file from gotFile(file), and dataFile which is loadJSON(file.data):

>>> file
>Object {
  >file: File,
  >_pInst: undefined,
  >type: "application", subtype: "json",
  >name: "poses_data_for_color.json", size: 33447,
  >data: "data:application/json;base64,eyJkYXRhIjpbeyJ4cyI6eyJpbnB1dDAiOjM4N…" }

>>> loadJSON(file.data)
>{}
  >data: Array(43) [ {…}, {…}, {…}, … ]
  ><prototype>: Object { … }

While the loadJSON(file.data) output looks pretty good, an object holding an array of {xs: {…}, ys: {…}}, I tried all kinds of extracting methods including dataFile['data'], dataFile[data], dataFile.data, dataFile[0]… and none of them worked.

The .json file downloaded from a previous model for loading.
The LINK to the p5 web editor of the second beta of the project.

Quick, Draw! Reflection

Quick, Draw! Dataset

Since all the images that are included in the dataset were submitted voluntarily by users and being low resolution and fidelity with minimal black lines, it’s hard for them to contain any confidential information. However, while the lines are so simple, what amazed me is that researchers can still identify tiny traces underlining the drawings including the directions of drawing a circle, which could disclose the nationality or the language the user is used to writing. The fact makes me doubt again about how could we protect our privacy in the era of AI and machine learning.

While the dataset is open-sourced, there are tasks and areas the data should not be used for. General speaking, according to Datasheet for Dataset, the use of AI in the areas of criminal justice, hiring, critical infrastructure, or finance is still doubtful. What’s more, the usages that could disclosure the personality or identity of individuals should also be limited whether or not there would be a potential harm to the person.

Feature Works

  1. Fix the bug of 'classification' model. As mentioned before, the classification brain is not working and currently I don't know if it's a bug from ml5 or my own code. (Fixed)

  2. A better output visualization, and interface design.