Simple-Human-in-the-Loop-ML-Interface

Description

A simple webpage-based human interface to facilitate human-in-the-loop machine learning research, which allows researchers to:

• remotely interact with (provide feedback) learning agent on server/clusters
• easily and efficiently annotate images
• replay any trajectory with specific frame time
• easily customize the interface to meet their needs (the interface is implemented with basic python Flask, HTML and JavaScript)

Getting Started

• Run the app.py as normal Flask program

python app.py

• Go to http://127.0.0.1:5000/
• Control and provide feedback through keyboard or buttons
• Create bounding boxes by simple click-hold-drag-release

Keyboard Input

You can define your own keyboard setting in function keyboardInputHandler in the file templates/index.html. Current setting is:

• A: positive feedback (+1)
• S: negative feedback (-1)
• D: no feedback (0)
• C: clear existing feedback
• >: next frame
• <: previous frame
• P: auto play/stop auto play

Some Examples of Customizing the Interface

• Object Tracking: object-tracking-example

Reference

• Simple Image Annotator: https://github.com/sgp715/simple_image_annotator
• Explanation Augmented Feedback in Human-in-the-Loop Reinforcement Learning: https://arxiv.org/abs/2006.14804