This repository contains the action recognition approach as presented in the Gimme Signals paper.

• A preprint can be found on arxiv.
• A video abstract is available on youtube.

In case the video does not play you can download it here

@inproceedings{Memmesheimer2020GSD, 
   author = {Memmesheimer, Raphael and Theisen, Nick and Paulus, Dietrich}, 
   title = {Gimme Signals: Discriminative signal encoding for multimodal activity recognition}, 
   year = {2020}, 
   booktitle = {2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)}, 
   address = {Las Vegas, NV, USA}, 
   publisher = {IEEE}, 
   doi = {10.1109/IROS45743.2020.9341699}, 
   isbn = {978-1-7281-6213-3}, 
 } 

• pytorch, torchvision, pytorch-lightning, hydra-core
• pip intall -r requirements.txt

The following command installs supported torch and torhvision versions in case you get an CUDA Kernel issue:

pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 -f https://download.pytorch.org/whl/torch_stable.html

Example code to generate representations for the NTU dataset:

python generate_representation_ntu.py <ntu_skeleton_dir> $DATASET_FOLDER <split>

where split is either cross_subject, cross_setup, one_shot

Representations must be placed inside a $DATASET_FOLDER that an environment variable points to.

We provide precalculated representations for intermediate result reproduction:

• NTU RGB+D 120 Cross Subject
• Simitate (MoCap)
• UTD-MHAD (Inertial and Skeleton)
• ARIL (Wi-Fi)

Example:

python train.py dataset=simitate model_name=efficientnet learning_rate=0.1 net="efficientnet"

Exemplary, this command trains using the simitate dataset.