This is the official repository of
MetaReconstruction: A Unified Framework for Reconstruction-based Video Anomaly Detection.
conda create -n meta_reconstruction python=3.10
conda activate meta_reconstruction
conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 -c pytorch
pip install -r requirements.txtDownload the CUHK Avenue, UCSD Ped2 and ShanghaiTech datasets and structure the data as follows:
dataset/
avenue
training
frames
01
.jpg
02
.jpg
...
testing
frames
01
.jpg
02
.jpg
...
avenue.mat
ped2
training
frames
01
.jpg
02
.jpg
...
testing
frames
01
.jpg
02
.jpg
...
avenue.mat
shanghaitech
training
videos
.avi
testing
frames
01_0014
.jpg
01_0015
.jpg
...
test_frame_mask
.npy
test_pixel_mask
.npy
To use our model, follow the code snippet below:
cd Reconstructed_based
# Train, Test and Demo 3D AutoEncoder
bash scripts/train_3dae.sh
bash scripts/eval_3dae.sh
bash demo.sh
# Train, Test and Demo STEAL
bash scripts/train_steal.sh
bash scripts/eval_steal.sh
bash demo.sh
# Train, Test and Demo MemAE
bash scripts/train_memae3d.sh
bash scripts/eval_memae3d.sh
bash demo.sh
# Train, Test and Demo Reconstruction MNAD
bash scripts/train_rmnad.sh
bash scripts/eval_rmnad.sh
bash scripts/demo.sh
# Train, Test and Demo Future Frame Prediction MNAD
bash scripts/train_pmnad.sh
bash scripts/eval_pmnad.sh
bash scripts/demo.shTBA.
If you find our work useful, please cite the following:
@misc{Chi2023,
author = {Chi Tran},
title = {MetaReconstruction: A Unified Framework for Reconstruction-based Video Anomaly Detection},
publisher = {GitHub},
booktitle = {GitHub repository},
howpublished = {https://github.com/IceIce1ce/MetaReconstruction},
year = {2023}
}
If you have any questions, feel free to contact Chi Tran
([email protected]).
Our framework is built using multiple open source, thanks for their great contributions.