Identifying and mapping urban Crime-Hotspots using a deep-learning model and sequential images of deviant locations (i.e., Location of reported deviant incidents such as crime and civil complaints).

• This repository contains the code for the DevianceNet applied to the deviance prediction.

• Sequential images
• Geo-tagged image files annotated with deviance class
• Dataset description

Please download the dataset at (https://drive.google.com/drive/folders/1ERFaC_6IseQgXDvs5_ep5at56b-b7Vmg?usp=sharing).

Deviance
├── Seoul
│    ├── Seoul_train_SEA 
│    │    ├── v_class1_g37.499,127.1251_c0
│    │    │     ├── frame000000.jpg
│    │    │     └── ...
│    │    └── ...
│    ├── Seoul_train_DIA
│    │    └── ...
│    ├── Seoul_test_SEA
│    │    └── ...
│    └── Seoul_test_DIA
│         └── ...
├── Busan
│    └── ...    
├── Incheon
│    └── ...    
├── Daejeon
│    └── ...    
├── Daegu
│    └── ...    
├── Newyork
│    └── ...    
└── Chicago
     └── ...   

The directory {City}_{Train/Test}_{SEA/DIA} is used for train/test corresponding to the {City}.

The files v_class{Number}_g{GPS}_c{Direction} is annotated with corresponding deviance class, GPS and viewpoint direction.

Note that '+'(ex. v_class2_g35.0911,129.0394_c1+) indicates the additional sets of sequential images in the GPS.

• Our implementation of DevianeNet (./models/devianceNet.py)
• Official implementation of DevianceNet (./main.py)
• Train & evaluation code for Deviance Dataset
• Code description

• Ubuntu 16.04
• python 3.8
• numpy>=1.18.5
• torch==1.7.0
• torchvision>=0.5.0
• pillow>= 8.0.1
• scikit-image
• tqdm
• sklearn
• pandas
• h5py
• matplotlib
• apex
• scipy>=1.4.1

1. Download the dataset and pretrained weights at (https://drive.google.com/drive/folders/1ERFaC_6IseQgXDvs5_ep5at56b-b7Vmg?usp=sharing).
2. Put the dataset into ./data/ folder
3. [Application & Evaluation] Put the pretrained weights into ./weight_file/ folder.

Use the 'main.py' to train/test our model.

# Train
python main.py --train_folder_directory {Dataset Directory} --SEA_folder_directory {Dataset Directory} --DIA_folder_directory {Dataset Directory} --experiment_description Train --batch_size 36 --num_threads 16 --classifier_type SEA_DIA
# Test SEA
python main.py --SEA_folder_directory {Dataset Directory} --experiment_description Test_SEA --batch_size 1 --classifier_type SEA_DIA --test_only --weight_load_pth ./weight_file/[AAAI2022-DevianceNet]SEA.pth --test_metric SEA
# Test DIA
python main.py --DIA_folder_directory {Dataset Directory} --experiment_description Test_DIA --batch_size 1 --classifier_type SEA_DIA --test_only --weight_load_pth ./weight_file/[AAAI2022-DevianceNet]DIA.pth --test_metric DIA
# Application
python main.py --SEA_folder_directory ./application --experiment_description Test_DIA --batch_size 1 --classifier_type SEA_DIA --test_only --weight_load_pth ./weight_file/[AAAI2022-DevianceNet]SEA.pth --test_metric SEA

If you find this code useful for your research, please cite our paper :)

@inproceedings{park2022deviance,
  title={DevianceNet: Learning to Predict Deviance from A Large-scale Geo-tagged Dataset},
  author={Park, Jin-Hwi and Park, Young-Jae and Lee, Junoh and Jeon, Hae-Gon},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2022}
}

Part of our code is borrowed from following implementations. We thank the authors for releasing their code and models.

• DELTAS: Depth Estimation by Learning Triangulation And densification of Sparse points (ECCV 2020) Code paper
• A Closer Look at Spatiotemporal Convolutions for Action Recognition (CVPR 2018) Code paper