What is it?
ConvnetClassifier is an openFrameworks application, part of the ml4a-ofx collection, which lets you train a classifier to recognize whatever objects, persons, postures, drawings and other visuals you show it through your webcam.
By default the app outputs OSC to localhost, port 8000, address “/classification”. This can be changed in ofApp.h.
As an alternative to using the GUI you can use the following key inputs:
- 1-9: set the class label
- s: save model
- l: load model
- t: train classifier
- c: clear examples
- r: toggle recording
Make sure your camera sees your intended visuals for class 1. Click [record] to toggle recording of training examples.
Set the Class Label slider to a different class and click [record] to record examples of a different class.
Repeat this for class 3 or however many distinct things you intend to classify. Remember to change the classes by pressing the numeric keys accordingly.
Click [train] to train the model.
Show the webcam different visuals and see what class the model predicts.
The video below takes you through the steps of the training process.
Setup and training considerations
ConvnetClassifier is a very versatile application that can be used in a wide variety of setups. However, to optimize correct classification, some simple rules of thumbs are:
Ensure variety in your training data: Say you want to classify whether someone is holding an object in their hands, try to record training samples that shows the person holding the object in different postures and positions, close to the camera, far from the camera, in both hands etc.
Keep your settings stable when using small training samples: Training in different sets of lighting conditions and hereafter attempting to classify in yet another environment might prove difficult. The same is true for camera positions and angles, backgrounds, etc. Unless you intentionally want a classifier that is can handle different settings (which would require more training examples), you can save yourself a lot of sweat by keeping your physical setup stable.
If the classification is not accurate enough, try to record more training examples. Alternatively you can try different classifier algorithms, which can be set in the setup() function of ofApp.cpp.