Skip to content
Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

HERA system

HERA (Heterogeneous Emotional Results Aggregator) is a three-level multimodal emotion detection system prepared to manage different emotion detectors in one place, detect emotions using said recognizers and aggregate results from each emotion detection service. logo16 The HERA system acts like a proxy of emotion recognizers: each emotion recognition service implements an interface in HERA, and the requests that would be send to said service are sent to HERA. HERA will then communicate with the corresponding service (being it a third-party service offered over the Internet or an API to access a sensor in a device), gathering the results and aggregate them on command.

Lanching HERA

Lanching HERA is quite simple. Once you've downloaded the repository, just run a

npm install

to install the node modules required and then a

npm start

to start the API. As an express app, it runs on port 3000 by default, but you can change this easily at the www file in /bin/. From this point on, you just send your requests to logo16 HERA in order to analyse data, aggregate results, etc. Since HERA works as an API Rest in a certain port, it's completely language agnostic. The snippets below show different examples of how to communicate with HERA using different programming languages. While these snippets are written in JavaScript, Python, Java and Ruby, you can use whatever programming language you want, as long as it supports communications via HTTP requests.

// Javascript (client)
    url : 'http://localhost:3000/api/v1/',
    type : 'POST',
    success : function( data ) {              
// JavaScript (Server)
	url: 'http://localhost:3000/api/v1/init',
	type: 'POST',
	body: {
		settingsFile: 'credentials.json'
	json: true
}, function( error, response, body ){
# Python
import requests
url = 'http://localhost:3000/api/v1/setup'
body = { 'delay': 3000 }
x = url, data = body )
// Java
RequestBody formBody = new FormBody.Builder()
	.add("mediaType", "image")
	.add("lookingFor", "['face']")
	.add("mediaPath", "C:\\Users\\user\\images\\image.jpg")

Request request = new Request.Builder()
# Ruby
require 'httparty'"http://localhost:3000/api/v1/results")

Unless you change the directory, the root of the request will always be /api/vX/, being X the number of the version you want to use. For example, if you want to use v1, an init request would have http://localhost:3000/api/v1/init as url.

Once the server is running, the workflow would be as follows:

  • Requesting a cookie. Call the root of the server to get an unique id. This id allows you to communicate with the rest of endpoints. Trying to communicate with these endpoints without including this unique id in the request will return an error (Session wasn't initialized. Send request to "/" first). This id is linked to an user object which will store the detectors' proxies, the data returned by these ones, etc.
  • Setting up the server. After getting the id, call the /init endpoint to create the detectors' proxies. This endpoint links a DetectorHandler to the user object, allowing you to request emotion recognition for media files and the aggregation of the consequent results.
  • Requesting a media analyse looking for emotions. Having your unique id, you will be able to request an emotion recognition over a media resource. Requests to this endpoint don't return the results to these analysis.
  • Aggregating results. Using the /results endpoint you can request the results to the server in several formats: aggregated, grouped by channel, results of a single detector, in PAD (translated) or RAW (as the detector produces them) format, etc.
  • (Optionally) Filtering detectors. After creating the detectors, you can optionally filter out detectors in the /setup endpoint based on several criteria, like the service latency, the kind of media it supports, etc.

See the next section to read more details about each endpoint.

Communicating with the API's endpoints

The core of the API is inside the routes/vX folders. Each vX folder (v1, v2, v3, etc.) contains an api.js file and a src folder. The api.js file contains the middleware to handle each endpoint of the API. This middleware is imported in the API entry point, (app.js file).

  • /

    • GET. Generates an user object with an unique id using uniqid and returns this id to the user in a cookie. CALLING THIS ENDPOINT IS MANDATORY TO BE ABLE TO USE THE API.
      • This endpoint has no parameters.
    • Return: User id stored in the cookie userId.
  • /init

    • POST. Inits detectors for an user, using setting information either sent in the request or stored in some setting file. This endpoint initializes each emotion detector and performs a benchmarking task (see DetectorHandler.prototype.addDetector) to test the state of the network and the detectors. See the bottom of the page to see an example of how this setting information must be specified. If none of these parameters are present in the request, no detector proxy will be created. If they are both specified, only settings will be used. REMEMBER: SEND JUST ONE OF THESE PARAMETERS.
      • settings {JSON}. JSON object following the aforementioned format.
      • settingsPath {string}. Path to file containing the setting information. Keep in mind that the route to the file starts at the root of the project(hera-system/...).
    • Return: Number of detectors initialized.
  • /analyse

    • POST. Endpoint used to analyse media files. The request contains the path to the file which holds the affective information, the type of information that the file contains and the kind of information that the API must look for.
      • mediaType {Array}: Kind of media which will be sent. Options can be "image", "video", "sound" and "text". However, any other type can be added to the API, since the media types that the API supports are specified on the /init endpoint, in the media attribute of each detector.
      • lookingFor {String}: Feature we want to analyse. Options can be "face", "voice", "signal" and "body". Again, the capabilities of the API regarding this parameter are specified on the /init endpoint, in the category of each detector.
      • mediaPath {String}: Path pointing to the file that must be analysed. This path can be either local or remote. Each detector must handle this path according to its characteristics. For instance, if you have a detector which can only analyse online resources, and you want it to analyse some local file (some image took with the webcam, a sound file recorded with the microphone), you'll have to upload somewhere inside the extractEmotions callback of said detector. For instance, in src/tools/upload-ftp.js you can find code to upload a file to a web server via FTP.
    • Return: HTTP status code and a message about the analysis process.


  • /results
    • POST. This endpoint retrieves a single PAD result, being this the product of aggregating all the results from every detector and channel.
      • channelsToMerge {Array|String}. Array of channels that should be merged. E.g., ["face", "voice"]. This parameters can also be a single string naming a channel ("face", "voice", "eda") or just "all".
      • localStrategy {String}. Strategy used to aggregate results locally. This strategy is used to by each detector of every channel present in channelsToMerge to aggregate its own results and then by every channel to aggregate those locally aggregated data.
      • globalStrategy {String}. Strategy used to aggregate the locally aggregated data. If any of these strategies doesn't exist, a default strategy (mean) is used.
    • Return: One triplet representing the dimensions Pleasure, Arousal and Dominance of all the aggregated data.
  • /setup
    • POST. This endpoint gives you another opportunity to customize the services to use. For instance, during the initialization in /init a benchmarking process is carried out. This process sets the value of the attributes realTime (boolean attribute which indicates if the service answers in real time) and delay (how many miliseconds does the service take to answer). This endpoint allows you to filter detectors that doesn't satisfy certain requirements, like working in real time, having a delay which is bigger than a fixed threshold, etc. /setup receives, in the request body, up to 3 parameters.
      • type {Array}: Array of the detector categories you want to keep. Detector categories which are not in this array will be deteled. An empty array deteles every category. E.g., ["face", "voice"].
      • realTime {Boolean}: boolean attribute representing the realTime attribute. Only detectors with a matching value of it will be kept.
      • delay: response time threshold. Detectors whose delay attribute is bigger than the one in the /setup request will be deleted.
    • Return: Number of detectors affected by this setup configuration.

Project structure

Here you can find how the project directory is organised. You can read the code documentation here, or you can access each specific method documentation through these links.











































In order to add support for new detectors, you just need to create a JavaScript file exporting the three following methods. Since HERA has been developed under an interface-based paradigm, these three methods are the only part you need to code to extend the API's functionality. Remember to export them using module.exports so the detector's methods can be accessed from outside the file.


This method takes care of any initialization tasks that your detector needs. For instance, if this detector is a proxy for a third-party service offered via Internet, you may need to get an auth token first. If this detector is a proxy for a bluetooth wristband, you may need to put your discovery and connection code in this method. If you don't need any initialization, just return a resolved promise (Promise.resolve). Important: this function must be async and always return a Promise.resolve, since the initializacion procress of every detector is async.


This is the main method of every detector, the one in charge of performing the actual emotion detection, being it forwarding some media resource to some API over the internet, reading RAW data from a sensor, etc. This method receives three parameters, namely context, media and callback. context is the environment from which the method is called, usually a Detector object; media is a path to the file which holds the affective information, if there is any. If there is no file (maybe the detector reads some RAW data from a certain port or socket in this method), this parameter will stay unused for the sake of the aforemention interface-based programming paradigm. Final parameter, callback, is an optional callback to handle the retrieved data, if there is any manipulation that has to be done.


In order to aggregate all the affective data, HERA needs to translate the data coming from each different service to the same format, that is, the PAD format. This way, each affective result coming from any detector is translated to a triplet of three numbers, being each one a value between -1 and 1 which stands for how negative or positive the expressed emotion is, how relaxed or excited the person is and how much passive or dominant does that person feel while feeling that emotion, respectively.

// E.g.
// Results coming from an API which analyses emotions present in a facial expression
const result = { happiness: 85.0, neutral: 1.2, surprise: 8.2, anger: 5.0 }
const padResult = translateToPAD(result);
// padResults = [0.75, -0.5,  -0.2]
// The image sent to the API presents a positive emotion, the subject seems rather calmed and shows a hint of modesty, which is translated into passivity.

// Results coming from a smart wristband which reads the acceleration of the movement of the hand, the heart rate of the person wearing it and their galvanic skin resistance (GSR.
const result1 = { acceleration: 10.54, heartrate: 96, gsr: 10 }
const padResult1 = translateToPAD(result);
// padResults1 = [0, 0.69, 0.5]
// The data gives no information about the positivity/negativity of the emotion, the subject is pretty active, based on their heart rate and galvanic skin resistance, so the emotion they're feeling is closer to excitement, and they are feeling pretty dominant, based in their movement (acceleration) and GSR.

It's completely in the developer's hands to decide how to do this transformation and how to consider the different parts of a result and how they affect the different dimensions of the PAD space.


settings.json sample file


	"detector-name-that-will-become-the-id": {
		"category": "face/voice/body/eda/...",
		"media": [ "types", "of", "media", "the service", "accepts", "for", "analysis", "such us", "image", "video", "text", "sound", "..." ],
		"realTime": boolean,
		"url": "https://endpoint-of-the-service-if-any",
		"otherOptions": {
			"key": "you may add all the data/optios you need in the otherOptions object"
		"callbacks": "./route/to/the/file/where/the/callbacks/are/defined.js"
	"detector-name-that-will-become-the-id-2": {
	"detector-name-that-will-become-the-id-3": {


Three-level multimodal emotion recognition framework to detect emotions combining different inputs with different formats.





No releases published


No packages published