Exploring Video-Based Driver Activity Recognition under Noisy Labels

PyTorch implementation of paper "Exploring Video-Based Driver Activity Recognition under Noisy Labels".

Dependencies

• Python, version = 3.10
• Ubuntu 16.04 Desktop
• PyTorch, version = 1.12.1
• CUDA, version = 11.6
• Anaconda3

Installation

Please find installation instructions for PyTorch and PySlowFast in INSTALL.md and follow the instructions in DATASET.md to prepare the datasets in Kinetics format.

Overview

• detectron2_repo/: Detectron2 repository
• slowfast/
  • config/: Config yaml files for different experiments
  • scripts/: Script ipynb files for dataset preparation and result visualization
  • slowfast/: Implementations
    • config/default.py: Default config
    • datasets/: Implementations of datasets
    • models/: Implementations of models
    • utils/: Helper functions
  • tools/: Scripts for training/validation/testing

Usage

In this repository, we provide the datast we used for this project. You should download the Drive&Act dataset and create the subset according to the csv files. The dataset should be put into the same folder of labels as the instructions in DATASET.md.

To generate noise labels, you can run the generate_noisy_label.ipynb in the script folder with any noise proportion.

Here is a training example:

python tools/run.py \
  --cfg configs/Kinetics/MViTv2_S_16x4.yaml \
  DATA.PATH_TO_DATA_DIR path_to_your_dataset \

To perform test, you can set the TRAIN.ENABLE to False, and do not forget to pass the path to the model you want to test to TEST.CHECKPOINT_FILE_PATH.

Acknowledgements

Great thanks for these open-source repositories:

• PySlowFast
• Drive&Act