This repository contains instructions to download the pre-processed data (from Ego4D, EK-100, and HOI4D datasets) used for the grasp affordance model training in "DEFT: Dexterous Fine-Tuning for Real-World Hand Policies." This data is modified from Ego4D, Epic-Kitchens, and HOI4D and pre-processed with the labels necessary for training.

Download data from here).

Alternatively, download directly to disk using gdown:

pip install gdown
gdown --folder 1E8RqVa8RDRlNJGX0FDt5aWV48ndo-XFJ

Unzip all data:

tar -xf deft-data-all/*.tar.gz -C deft-data-all
rm deft-data-all/*.tar.gz

This data is modified from the Ego4D, Epic-Kitchens 100, and HOI4D datasets. We used a subset of the images from the videos recorded those datasets, detected the affordances, and provided labels for model training. Both Epic Kitchena and HOI4D are licensed under CC BY-NC 4.0, and Ego4D's license is here.

This dataset is licensed under CC BY-NC 4.0.

When using this dataset, please reference:

@article{kannan2023deft,
         title={DEFT: Dexterous Fine-Tuning for Real-World Hand Policies},
         author={Kannan, Aditya, and Shaw, Kenneth and Bahl, Shikhar, and Mannam, Pragna and Pathak, Deepak},
         journal={Conference on Robot Learning (CoRL)},
         year={2023}}