The work is developed in collaboration with bach05 and albigotta. Please cite these papers in your publications if FSG-Net helps your research.
@inproceedings{barcellona2023fsg,
title={FSG-Net: a Deep Learning model for Semantic Robot Grasping through Few-Shot Learning},
author={Barcellona, Leonardo and Bacchin, Alberto and Gottardi, Alberto and Menegatti, Emanuele and Ghidoni, Stefano},
booktitle={2023 IEEE International Conference on Robotics and Automation (ICRA)},
pages={1793--1799},
year={2023},
organization={IEEE}
}
The repository is under construction to provide a user-friendly interface and a more readable code. The draft code we developed is already available.
pip install -r requirements.txt
Pre-trained models are available, please refer to the following list:
- Few-Shot Module: LINK, destination folder:
./output/models/fewshot_modules
- Backbone: LINK, destination folder:
./output/models/backbones
- Quality Head: LINK, destination folder:
./output/models/heads
- Angle Head: LINK, destination folder:
./output/models/heads
- Width Head: LINK, destination folder:
./output/models/heads
folder.
The training process (e.g. training hyperparameters, paths, ecc...) can be configured through ./config/generic.yaml
files.
[DETAILED EXPLANTION]
python3 train.py
[ADD THE CODE FOR TESTING]