The official PyTorch implementation of the paper "Normalizing Flows for Human Pose Anomaly Detection".

If you find this useful, please cite this work as follows:

@article{hirschorn2022human,
  title = {Normalizing Flows for Human Pose Anomaly Detection},
  author = {Hirschorn, Or and Avidan, Shai},
  journal={arXiv preprint arXiv:2211.10946},
  year = {2022},
}

This code was tested on Ubuntu 20.04.4 LTS and requires:

• Python 3.8
• conda3 or miniconda3
• CUDA capable GPU (one is enough)

git clone https://github.com/orhir/STG-NF
cd STG-NF

# Conda environment setup
conda env create -f environment.yml
conda activate STG-NF

.
├── checkpoints
├── data
│   ├── ShanghaiTech
│   │   ├── gt
│   │   │   └── test_frame_mask
│   │   └── pose
│   │       ├── test
│   │       └── train
│   └── UBnormal
│       ├── gt
│       ├── pose
│       │   ├── abnormal_train
│       │   ├── test
│       │   ├── train
│       │   └── validation
│       └── videos
├── models
│   └── STG_NF
└── utils

Data folder, including extracted poses and GT, can be downloaded using the link.

The data directory holds pose graphs and ground truth vectors for the datasets. A path for the directory may be configured using the arguments:

--vid_path_train
--vid_path_test
--pose_path_train
--pose_path_train_abnormal
--pose_path_test

We provide a script for creating JSON files in the accepted format using AlphaPose. Please download into pretrained_models folders fast_421_res152_256x192.pth It is also recommended to use the YOLOX-X detector, which can be downloaded from the AlphaPose repository. Use the flag --video for video folder, otherwise assumes a folder of JPG/PNG images for each video. python gen_data.py --alphapose_dir /path/to/AlphaPoseFloder/ --dir /input/dir/ --outdir /output/dir/ [--video]

Training and Evaluating is run using:

python train_eval.py --dataset [ShanghaiTech\UBnormal]

Evaluation of our pretrained model can be done using:

ShanghaiTech/ShanghaiTech-HR:

python train_eval.py --dataset [ShanghaiTech/ShanghaiTech-HR] --checkpoint checkpoints/ShanghaiTech_85_9.tar

Unsupervised UBnormal

python train_eval.py --dataset UBnormal --seg_len 16 --checkpoint checkpoints/UBnormal_unsupervised_71_8.tar 

Supervised UBnormal

python train_eval.py --dataset UBnormal --seg_len 16 --R 10 --checkpoint checkpoints/UBnormal_supervised_79_2.tar

Our code is based on code from:

• Graph Embedded Pose Clustering for Anomaly Detection
• Glow

This code is distributed under a Creative Commons LICENSE.

Note that our code depends on other libraries and uses datasets that each have their own respective licenses that must also be followed.