Few-Shot Console

This repository contains a simple inference script for the models of the paper XXX (currently under blind review, contact @steb6 for a copy) and a GUI that allow to easily collect new demonstrations and to modify known classes.

Setup

Clone the few-shot-console repository:

git clone https://github.com/hsp-iit/few-shot-console.git

Install Conda, then create a Conda environment with:

conda env create -f environment.yml

Download the model checkpoints depending on your target settings:

• Table top: download the checkpoint from SSv2
• Human-Robot Interaction: download the checkpoint from NTURGBD120

Launch

If needed, adjust the setting in config.json. Within the conda environment activated, launch

python demo.py

Note that the demo.py script search for available cameras with OpenCV and then select the last one as input source. Modify this behaviour if needed.

There are two Support Set examples that can be loaded:

• coffee: table-top, use SSv2 checkpoint for this scenario
• hri: human-robot-interaction, use NTURGBD120 for this scenario