TEZARNet - TEmporal Zero-shot Activity Recognition Network
Abstract
Most existing Zero-Shot Human Activity Recognition (ZS-HAR) methods based on Inertial Measurement Unit (IMU) data rely on attributes or word embeddings of class labels as auxiliary data to relate the seen and unseen classes. However, defining attributes requires expert knowledge, and both attributes and word embeddings lack motion-specific information. On the other hand, videos depicting various human activities are readily available and contain valuable information for ZS-HAR based on inertial sensor data. This paper proposes a new IMU-data-based ZS-HAR model using videos as auxiliary data. In contrast to the current work, we employ a Bidirectional Long-Short Term based IMU encoder to exploit the temporal information. The proposed model outperforms the state-of-the-art accuracy by 4.7%, 7.8%, 3.7%, and 9.3% for benchmark datasets PAMAP2, DaLiAc, UTD-MHAD, and MHEALTH, respectively.
Overview of TEZARNET Training Phase
Preparation
clone the code base
setup libraries
python -m pip install requirements.txt
setup Git LFSgit lfs install
git lfs track
download imu datasets 1. download datasets directly from the repo 2. download datasets from original source download video datasets corresponding video datasets are link in here.
Execution
Training
- for PAMAP2 dataset with default configs
python main.py --IMU_data_path ./data_path --I3D_data_path ./data_path
- for DaLiAc dataset
python main.py --IMU_data_path ./data_path --I3D_data_path ./data_path --datasets daliac --d_model 224
Citation
@inproceedings{deelaka2023texarnet,
title={TEZARNet - TEmporal Zero-shot Activity Recognition Network},
author={Deelaka, Pathirage N. and Y. De Silva, Devin and Wickramanayake, Sandareka and Meedeniya, Dulani and Rasnayaka, Sanka},
booktitle={International Conference on Neural Information Processing},
year={2023},
organization={Springer}
}