Skip to content

Latest commit

 

History

History
99 lines (75 loc) · 5.81 KB

README.md

File metadata and controls

99 lines (75 loc) · 5.81 KB
Raw

Crowded scenes Ensemble classification

This is the companion repository for our paper titled "Classification ensembliste de vidéos de mouvements de foule" accepted for presentation at the french national conference ORASIS 2021 and our paper titled "Ensemble classification of video-recorded crowd movements" accepted for presentation in the ISPA 2021 IEEE conference. The project is about the Ensemble classification of 10 crowd movements illustrated in the Crowd-11 dataset. The 11th class is intended for empty scenes.

Four different architectures are employed for the Ensemble classification:

@inproceedings{tran2015learning,
  title={Learning spatiotemporal features with 3d convolutional networks},
  author={Tran, Du and Bourdev, Lubomir and Fergus, Rob and Torresani, Lorenzo and Paluri, Manohar},
  booktitle={Proceedings of the IEEE international conference on computer vision},
  pages={4489--4497},
  year={2015}
}

The implementation of C3D in Keras was forked from here.

@inproceedings{carreira2017quo,
  title={Quo vadis, action recognition? a new model and the kinetics dataset},
  author={Carreira, Joao and Zisserman, Andrew},
  booktitle={proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={6299--6308},
  year={2017}
}

The implementation of I3D in Keras was forked from here. According to the authors of the keras-kinetics-i3d repository, the weights of I3D, that we provide in the Data/ folder, were obtained from here and are under Apache-2.0 License.

@inproceedings{hara2017learning,
  title={Learning spatio-temporal features with 3d residual networks for action recognition},
  author={Hara, Kensho and Kataoka, Hirokatsu and Satoh, Yutaka},
  booktitle={Proceedings of the IEEE International Conference on Computer Vision Workshops},
  pages={3154--3160},
  year={2017}
}

We used the keras implementation of the R3D architecture that can be found here.

Requirements

Refer to the requirements.txt file to install the required versions of tensorflow-gpu, Keras, Opencv, Numpy. You may also need to install Matplotlib, Pandas, Scikit-learn, Scikit-videos, Scikit-image.

Download the Crowd-11 dataset

Instructions on how to get the Crowd-11 dataset may be found in the following workshop paper : Crowd-11: A Dataset for Fine Grained Crowd Behaviour Analysis

@inproceedings{dupont2017crowd,
  title={Crowd-11: A dataset for fine grained crowd behaviour analysis},
  author={Dupont, Camille and Tobias, Luis and Luvison, Bertrand},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops},
  pages={9--16},
  year={2017}
}

Usage

After downloading the Crowd-11 dataset, extract the optical flow and put both of the flow and the rgb clips inside the subfolders of the Data/Crowd-11/ folder, like this:

-- Data/
    -- Crowd-11/
        -- rgb/
        -- flow/

The optical flow is extracted using the following forked and updated project: https://github.com/MounirB/py-denseflow

When you obtain the Crowd-11 dataset, it is crucial to include the preprocessing.csv file into the Data folder before generating the k Folds.

Generate k Folds

Before creating homogeneous Ensembles based on the dataset folds, you should split the dataset into several folds. To do so, launch the following script generate_folds.sh that will run the generate_folds.py script. By default, the number of folds K is set to 5. In the main program of the generate_folds.py script, you can find that the program relies on preprocessing.csv spreadsheet. Remove, if needed, the missing clips paths of Crowd-11 form preprocessing.csv.

Data augmentation

To augment the dataset folds, use the augment_dataset.sh script. In this script, you should specify the augmentation frequency --augmentation_frequency. The augmentation script augment_dataset.py is forked from this project: https://github.com/okankop/vidaug

Train Ensembles

To train an Ensemble of homogeneous models, run the following script launch_train_ensemble.sh. The options in the script will guide you to choose the unique architecture of the models of the Ensemble.

Evaluate Ensembles of Homogeneous models

To evaluate an Ensemble of homogeneous models, run the following script launch_evaluate_ensemble.sh. The options in the script will guide you to choose the pre-trained Ensemble to evaluate.

Evaluate Global Ensembles of heterogeneous Ensembles of homogeneous models

To evaluate a Global Ensemble of heterogeneous Ensembles of homogeneous models or evaluate different combinations, run once again the following script launch_evaluate_ensemble.sh.