TrashNet takes a multi-pronged approach to identifying trash, building two classes of models. One of these(EfficientDet) is designed for throughput, enabling parallel inference. The other class of models, using the SSD MobileNet v2 architecture is designed for edge devices, allowing you to do on-premise inference.
I also wrote a Streamlit app to help visualize the predictions from the model, allowing you to inspect your model performance. As I’ve built TrashNet, this has been especially useful in letting me know what data my model needs more learning on. The webapp is on course to be deployed internally at CleanRobotics by mid July 2020.
For a writeup on the project, click here.
For info on setting up TrashNet and other tutorials, head to the documentation here.