What is it?

ConvnetOSC is an openFrameworks application, part of the ml4a-ofx collection, which sends ConvNet activations as a 4096-bit input vector over OSC. Used in combination with Wekinator ConvnetOSC lets you train a classifier to recognise whatever objects, persons, postures, drawings and other visuals you show it through your webcam. It can also be used for regression.

ConvnetOSC gif

ConvnetOSC has been used by Andreas Refsgaard in his installation IS IT FUNKY? which tries distinguish funky images from boring ones after being trained on 15.000 (highly subjective) images from Google searches.

Training instructions

The video below takes you through the steps of the training process.

To change the OSC host, port, and address, edit the settings.xml file in the data folder. By default, it is set to localhost, port 6448, address “/wek/inputs”, which works out of the box with Wekinator.

Setup and training considerations

ConvnetOSC is a very versatile application that can be used in a wide variety of setups. However, to optimise correct classification/regression, some simple rules of thumbs are:

How does a convnet work?