Protest Activity Detection and Perceived Violence Estimation from Social Media Images

Implementation of the model used in the paper Protest Activity Detection and Perceived Violence Estimation from Social Media Images (ACM Multimedia 2017) [arxiv] by Donghyeon Won, Zachary C. Steinert-Threlkeld, Jungseock Joo.

Requirements

Pytorch
NumPy
pandas
scikit-learn

Usage

Training

python train.py --data_dir UCLA-protest/ --batch_size 32 --lr 0.002 --print_freq 100 --epochs 100 --cuda

Evaluation

python pred.py --img_dir path/to/some/image/directory/ --output_csvpath result.csv --model model_best.pth.tar --cuda

UCLA Protest Image Dataset

You will need to download our UCLA Protest Image Dataset to train the model. Please e-mail me if you want to download our dataset!

Dataset Statistics

# of images: 40,764
# of protest images: 11,659

Protest & Visual Attributes

Fields	Protest	Sign	Photo	Fire	Police	Children	Group>20	Group>100	Flag	Night	Shouting
# of Images	11,659	9,669	428	667	792	347	8,510	2,939	970	987	548
Positive Rate	0.286	0.829	0.037	0.057	0.068	0.030	0.730	0.252	0.083	0.085	0.047

Violence

Mean	Median	STD
0.365	0.352	0.144

Model

Architecture

We fine-tuned ImageNet pretrained ResNet50 to our data. You can download the model I trained from this Dropbox link.

Performance

Protest	Sign	Photo

Fire	Police	Children

Group>20	Group>100	Flag

Night	Shouting	Violence