This project is just still prototype, very simple real-time pose classification, yet this project worked perfectly fine.
Built using Javascript with the help of ml5js and p5js library.
Read this blog for more information about real-time pose estimation using javascript.
- You can see all dependencies inside the
<head>tag. - Live Server Plugin.
- Clone this repository using
git clonecommand (or just download thezipversion). - Extract the cloned project.
- Enter the directory.
$ cd $home/this-cloned-project
- Open the project using VSCode or anything you like as long as that code editor has Live Server Plugin enabled, otherwise this project won't work.
-
Go to Home Page or Collect Pose Page first. There's menu on top of it, you can choose which session you want.
Below are the detailed information about each page :
-
Collect Pose Page
- After the webcam loaded you just press keyboard (except for 'S' key) to start collect the data.
- After you pressed the key, the state will switch to 'Collecting' and then you can start collect the pose until the state switch to 'Waiting' (it's about 15s of collecting data). You can repeat this process as much as you like.
- Press the 'S' key to save the data. Rename the data as
data.jsonand save to$this-project/data/directory
-
Train Data Page
- This page will load the saved data from Collect Pose Page, you can see or change the loaded data name inside
$this-project/js/dataTraining.js.function setup() { ... brain.loadData('data/data.json', dataReady); ... }
- The data training process may take a while depends on epoch's value.
- After the data training process done it will automatically download 3 files, save those files to
$this-project/data/directory with default name.
- This page will load the saved data from Collect Pose Page, you can see or change the loaded data name inside
-
Classify Pose Page
- This page will load saved data from Train Data page, you can see or change the loaded data name inside
$this-project/js/dataClassification.js.function setup() { ... const modelInfo = { model: 'data/model.json', metadata: 'data/model_meta.json', weights: 'data/model.weights.bin', }; ... }
- After the camera loaded you can start posing, and this page will try to classify your pose based on the collected pose from Collect Pose Page.
- This page will load saved data from Train Data page, you can see or change the loaded data name inside
To configure how much pose you want to save, you can customize that value inside setup() function inside dataClassification.js and dataCollection.js
function setup() {
...
let options = {
inputs: 34, // How much keypoints as input, 17 keypoints consist of x and y pairs so it's 34 coordinates
output: 4, // How much pose you want to save
task: 'classification',
debug: true
}
...
}List of 'things' that may affect the Collecting Data Process :
- Position
- Lighting
- Outfit
- Background
As you can see POSEFICATION is under MIT License.