This is the github repository containing the code for the paper "Generalizing Hand Segmentation in Egocentric Videos with Uncertainty-Guided Model Adaptation" by Minjie Cai, Feng Lu and Yoichi Sato.

Currently, the code is under update.

The code is tested to work correctly with:

• Ubuntu 18.04
• GPU environment
• Anaconda Python 3.7
• Pytorch v0.4.0
• NumPy
• OpenCV
• tqdm

It may also work with other versions of Ubuntu, Python 3 and Pytorch, but we haven't checked.

We use several different datasets for model adaptation (or domain adaptation), including: EGTEA, GTEA, EDSH2, EDSH-Kitchen, UTG, YHG, and EgoHands. The UTG and Yale_Human_Grasp datasets are prepared by the paper and included in the repository. Other datasets are public and please download by yourself and arrange them in filefolders like UTG.

You can download the hand segmentation model (bayes_rf101_egtea_00040.pth.tar) pre-trained on EGTEA dataset from BaiduDrive code: rvch
Please put the model inside the filefolder according to the filepath of the code.
To generate the pre-trained model for yourself, after preparing the source domain data (EGTEA dataset), run

python train.py --dataset name_of_source_dataset --batchSize your_batchsize_for_training

You can also find the source domain data from the above BaiduDrive.

To run the complete experiment, after preparing the data, run

python train_bayesian_adapt.py --dataset name_of_target_dataset --batchSize your_batchsize_for_adaptation

This would adapt the pre-trained segmentation model to the target dataset.

Please cite the following paper if you feel this repository useful.

@inproceedings{cai2020generalizing,
  author    = {Minjie Cai and
               Feng Lu and
               Yoichi Sato},
  title     = {Generalizing Hand Segmentation in Egocentric Videos with Uncertainty-Guided Model Adaptation},
  booktitle   = {IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2020},
}

For any question, please contact

Minjie Cai: [email protected]