GROWL_public

Repository holding the implementation for GROWL - Group Detection With Link Prediction

References

When using this code and/or the RICA dataset, please cite the following works:

RICA: Robocentric Indoor Crowd Analysis Dataset

@article{schmuck2020rica,
  title={RICA: Robocentric Indoor Crowd Analysis Dataset},
  author={Schmuck, Viktor and Celiktutan, Oya},
  journal={IMU},
  volume={127},
  number={74,234},
  pages={31--172},
  year={2020}
}

GROWL: Group Detection With Link Prediction

@inproceedings{schmuck2021growl,
  title={GROWL: Group Detection With Link Prediction},
  author={Schmuck, Viktor and Celiktutan, Oya},
  booktitle={2021 16th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2021)},
  pages={1--8},
  year={2021},
  organization={IEEE}
}

Installation

• Download or clone the repository
• Set up an environment with python 3.9
• Install the following packages (indicating latest tested compatible versions):
  • pandas==1.3.5
  • matplotlib==3.5.1
  • scipy==1.7.3
  • numpy==1.22.1
  • dgl==0.6.1
  • torch==1.10.1
  • networkx==2.6.3
  • scikit-learn==1.0.2

Quick installation commands if you are using conda

conda create --name growl_pub python=3.9
pip install pandas matplotlib scipy numpy dgl torch networkx scikit-learn

Usage

For multiple tests, run retest_growl.py, changing the parameters in lines 20-33 as you see fit. These are described below:

• feats: number of features to create during embedding
• epochs: number of epoch to train for
• balanced: whether only training runs with balanced negative and positive edge samples should be executed. Set to balanced='' if not.
• string_end: how the output csv filename should end, usually a date
• ablation: if an ablation test should be executed. Choose a corresponding index from the ablations list.
• test_on: choose which dataset the tests should be executed on. Choose a corresponding index from the test_on_options list.
• test_iters: how many iterations of the test should be executed.

---

It's also possible to execute individual tests by calling (e.g.):

python growl_test_code.py -f 20 -e 100 -s 20220121 -t all_salsa -a no -b

Command line parameter usage:

usage: growl_test_code.py [-h] [-f FEATS] [-e EPOCHS] [-s STRING_END] [-t {ps,cpp,all_salsa,rica,rica_yolo}] [-a {no,ori,edg}] [-b] [-p] [--yolo_acc] [--yolo_to_gt]

To perform tests with GROWL, decide which data to test on and whether this is an ablation test.

optional arguments:
  -h, --help            show this help message and exit
  -f FEATS, --feats FEATS
                        Sets the number of features to create an embedding for.
  -e EPOCHS, --epochs EPOCHS
                        Sets the numebr of epochs to train GROWL for.
  -s STRING_END, --string_end STRING_END
                        The postfix of the results filename.
  -t {ps,cpp,all_salsa,rica,rica_yolo}, --test {ps,cpp,all_salsa,rica,rica_yolo}
                        Which dataset to test on.
  -a {no,ori,edg}, --ablation {no,ori,edg}
                        Is this a test involving ablation?
  -b, --balance_samples
                        Enable sample balancing to discard runs if there is too big of an imbalance between positive and negative edge counts.
  -p, --plot            Enable plotting the graphs produced during the tests.
  --yolo_acc            Print YOLOv4 detection accuracy compared to RICA's ground truth. Only has an effect if test=rica_yolo.
  --yolo_to_gt          When testing on YOLOv4 detections, calculate F1-scores of detected groups compared to the GT detections.