Source code for paper "Multistream Temporal Convolutional Network for Correct/Incorrect Patient Transfer Action Detection Using Body Sensor Network" on public action detection dataset C-MHAD
To give multimodal signals full play in fine-grained action detection, we propose a novel temporal convolutional network by designing a channel attention-based multistream structure. Compared to the state-of-the-art methods, our method achieves the best performance, using only inertial data, with F1 score of 95.3% for the action set of transition movements and 98.5% for the action set of smart TV gestures. We surpass the benchmark performance (F1 score of 78.8% for transition movements and 81.8% for smart TV gestures) using fused data, i.e., both inertial and video signals, by a large margin.
Install the dependent packages:
conda create -n cad python=3.8
conda activate cad
pip install -r requirements.txt
Download inertial data of C-MHAD directly from here (or from C-MHAD project page).
Train on Transition Movements part of C-MHAD:
python main.py --dataset CMHAD_Transition
Train on Smart TV Gesture part of C-MHAD:
python main.py --dataset CMHAD_Gesture
Please download checkpoints and unzip it under the main directory.
Run the pretrained model on Transition Movements part of C-MHAD:
python main.py --dataset CMHAD_Transition --test_only --test_checkpoint ./checkpoints/MSSTCN_CMHAD_Transition.tar
Run the pretrained model on Smart TV Gesture part part of C-MHAD:
python main.py --dataset CMHAD_Gesture --test_only --test_checkpoint ./checkpoints/MSSTCN_CMHAD_Gesture.tar
If you find our code is useful for you, please consider citing:
@article{zhong2021multistream,
title={Multistream Temporal Convolutional Network for Correct/Incorrect Patient Transfer Action Detection Using Body Sensor Network},
author={Zhong, Zhihang and Lin, Chingszu and Kanai-Pak, Masako and Maeda, Jukai and Kitajima, Yasuko and Nakamura, Mitsuhiro and Kuwahara, Noriaki and Ogata, Taiki and Ota, Jun},
journal={IEEE Internet of Things Journal},
year={2021},
publisher={IEEE}
}