srikarym / cctv-gun Goto Github PK

This repo contains training and evaluation code of CCTV-GUN benchmark. It uses mmdetection to train object detection models. Our arXiv submission can be found here.

We follow the installation instructions in the mmdetection documentation here. Specifically, our code requires mmcls=0.25.0, mmcv-full=1.7.0 and torch=1.13.0.

The output of conda env export > env.yml can be found in env.yml. It can be used to create a conda virtual environment with

conda env create  -f env.yml
conda activate env_cc
pip install openmim
mim install mmcv-full==1.7.0
pip install -e . 

We use images from three datasets :

1. Monash Gun dataset (MGD) ¹
2. US Real-time Gun detection dataset (USRT) ²
3. UCF Crime scene dataset (UCF) ³

Instructions on how to download these datasets can be found in dataset_instructions.md .

We perform two kinds of assessment : Intra-dataset and Cross-dataset (See paper for more details). We train five detection models :

• Faster R-CNN ⁴
• Swin-T ⁵
• Deformable DETR ⁶
• DetectoRS ⁷
• ConvNeXt-T ⁸

All of the above datasets consists of two classes : Person (class 0) and Handgun (class 1). To train a detection model on this dataset, run

python tools/train.py --config <path/to/model/config.py> --dataset-config <path/to/dataset/config.py> <extra_args>

• Model config files link

• Dataset config files link

• Trained models link

To adjust the training batch size

<base_command> --cfg-options data.samples_per_gpu=<batch-size>

Using weights and biases to log metrics: After you create an account in wandb, change entity and project in train.py to your wandb username and project name. Then

<base_command> --use-wandb --wandb-name <name-of-the-experiment>

Train a Swin-T on MGD (Intra-dataset)

python tools/train.py --config configs/gun_detection/swin_transformer.py --dataset-config configs/_base_/datasets/gun_detection/mgd.py --cfg-options data.samples_per_gpu=6

Train a detectoRS model on MGD+USRT (Inter-dataset)

python tools/train.py --config configs/gun_detection/detectors.py --dataset-config configs/_base_/datasets/gun_detection/mgd_usrt.py --cfg-options data.samples_per_gpu=4

Fine-tune MGD+USRT trained Deformable-DETR model on UCF (Inter-dataset)

python tools/train.py --config configs/gun_detection/deformable_detr.py --dataset-config configs/_base_/datasets/gun_detection/ucf.py --cfg-options data.samples_per_gpu=6 --load-from <path/to/trained/model.pth>

To evaluate a trained model, run

python tools/test.py --config <path/to/model/config.py> --dataset-config <path/to/dataset/config.py> --checkpoint <path/to/trained/model> --work-dir <path/to/save/test/scores> --eval bbox

Evaluate a faster-rcnn trained on MGD on MGD's test set (Intra-dataset)

python tools/test.py --config configs/gun_detection/faster_rcnn.py --dataset-config configs/_base_/datasets/gun_detection/mgd.py --checkpoint <path/to/mgd/trained/model.pth> --work-dir <path/to/save/test/scores> --eval bbox

Evaluate a ConvNeXt trained on MGD+USRT on the entirety of UCF (Inter-dataset)

python tools/test.py --config configs/gun_detection/convnext.py --dataset-config configs/_base/datasets/gun_detection/ucf_test_full.py --checkpoint <path/to/mgd+usrt/trained/model.pth> --work-dir <path/to/save/test/scores> --eval bbox

To save the bounding box predictions on test set , add --save-path <path/to/output/folder> to the above command.

Please consider citing the following references in your publications if it helps your research

@misc{yellapragada2023cctvgun,
      title={CCTV-Gun: Benchmarking Handgun Detection in CCTV Images}, 
      author={Srikar Yellapragada and Zhenghong Li and Kevin Bhadresh Doshi and Purva Makarand Mhasakar and Heng Fan and Jie Wei and Erik Blasch and Haibin Ling},
      year={2023},
      eprint={2303.10703},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

@article{lim2021deep,
  title={Deep multi-level feature pyramids: Application for non-canonical firearm detection in video surveillance},
  author={Lim, JunYi and Al Jobayer, Md Istiaque and Baskaran, Vishnu Monn and Lim, Joanne MunYee and See, John and Wong, KokSheik},
  journal={Engineering applications of artificial intelligence},
  volume={97},
  pages={104094},
  year={2021},
  publisher={Elsevier}
}

@article{gonzalez2020real,
  title={Real-time gun detection in CCTV: An open problem},
  author={Gonz{\'a}lez, Jose L Salazar and Zaccaro, Carlos and {\'A}lvarez-Garc{\'\i}a, Juan A and Morillo, Luis M Soria and Caparrini, Fernando Sancho},
  journal={Neural networks},
  volume={132},
  pages={297--308},
  year={2020},
  publisher={Elsevier}
}

@inproceedings{sultani2018real,
  title={Real-world anomaly detection in surveillance videos},
  author={Sultani, Waqas and Chen, Chen and Shah, Mubarak},
  booktitle={Proceedings of the IEEE conference on computer vision and pattern recognition},
  pages={6479--6488},
  year={2018}
}