You may notice that your package detector is sending you notifications when things that aren't packages are in its view (e.g. people, birds, etc.).

To improve our model, we'll use continuous learning by taking pictures of things that confuse our model, and then retrain the model using these new images.

Collect some other objects that aren't packages that might be by your front door. Wait for the model to make a prediction, then push and hold the joy stick up (towards the LEDs) on the BrainCraft if the label you see is correct, the LEDs will blink green when the picture is saved. If the label is wrong, push and hold the joy stick down (away from the LEDs) the LEDs will blink red when the picture is saved. 

This will save the images into a structed dataset folder named retraining_data.

Take 10-20 pictures of each object using the joystick on the Pi. 

Open an FTP Connection to the Pi and copy the retraining_data folder to your computer.

Open the Package Detector project in Lobe.

Import the structured dataset you created on your Pi.

Any images saved in the top level directory, retraining_data in this case, will need to be labeled in Lobe. You can put these photos in the "no package" category or, if you want a more detailed package detector, you can create new categories!

When you're done labeling, export your updated model and deploy it to the Pi. Follow instructions in "Export your Model". 

This guide was first published on Mar 30, 2021. It was last updated on 2021-03-30 15:53:33 -0400.

This page (Continuous Learning) was last updated on Jun 16, 2021.

Text editor powered by tinymce.