Source-free Video Domain Adaptation by Learning from Noisy Labels

This is the official code repository for "Source-free Video Domain Adaptation by Learning from Noisy Labels", Arxiv'24. An initial version of this work is published at ICVGIP'22.

Requirements

To install dependencies, please use the following command -

conda env create -f environment.yml

Training:

To reproduce the results reported in the paper, please follow the steps given below -

Step 1: Prepare the dataset

data
├── flow
├── rgb
|   ├── ucf101
|   |   ├──  v_YoYo_g25_c05
|   |   ├──  ...
|   ├── hmdb51

Step 2: Source-only Pre-training

You may need to adjust the data path in the script

bash scripts/source_only_train.sh ucf101 hmdb51 Joint

Step 2: Source-only Pre-training

bash scripts/generate_pseudo_labels.sh ucf101 hmdb51 Joint 12

Step 2: Adaptation Training

To run the CleanAdapt, assuming \tau = 0.5 -

bash scripts/adaptation_uh.sh ucf101 hmdb51 Joint 0.5

To run the CleanAdapt + TS, assuming \tau = 0.5 -

bash scripts/adaptation_uh_ema.sh ucf101 hmdb51 Joint 0.5

Please check the parse_args.py for more details on the argumments.

Citation:

Please consider citing the following work if you make use of this repository:

@inproceedings{dasgupta2024source,
  title={Source-free Video Domain Adaptation by Learning from Noisy Labels},
  author={Dasgupta, Avijit and Jawahar, CV and Alahari, Karteek},
  booktitle={Arxiv},
  year={2024}

@inproceedings{dasgupta2022overcoming,
  title={Overcoming Label Noise for Source-free Unsupervised Video Domain Adaptation},
  author={Dasgupta, Avijit and Jawahar, CV and Alahari, Karteek},
  booktitle={ICVGIP},
  year={2022}
}

Contact

In case of any issues, feel free to create a pull request. Or reach out to Avijit Dasgupta.