Latent Space Autoregression for Novelty Detection

This repository contains Pytorch code to replicate experiments in the CVPR19 paper "Latent Space Autoregression for Novelty Detection".

Please cite with the following BibTeX:

@inproceedings{abati2019latent,
  title={{Latent Space Autoregression for Novelty Detection}},
  author={Abati, Davide and Porrello, Angelo and Calderara, Simone and Cucchiara, Rita},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision and Pattern Recognition},
  year={2019}
}

Specifically, performs:

• one class classification on MNIST.
• one class classification on CIFAR-10.
• video anomaly detection on UCSD Ped2.
• video anomaly detection on ShanghaiTech.

0 - Clone this repo

First things first, clone this repository locally via git.

git clone https://github.com/cvpr19-858/novelty-detection.git
cd novelty-detection

1 - Environment

This code runs on Python 3.6. The easiest way to set up the environment is via pip and the file requirements.txt:

pip install -r requirements.txt

2 - Datasets

MNIST and CIFAR-10 will be downloaded for you by torchvision.

You still need to download UCSD Ped and ShanghaiTech. After download, please unpack them into the data folder as follows

tar -xzvf <path-to-UCSD_Anomaly_Dataset.tar.gz> -C data
tar -xzvf <path-to-shanghaitech.tar.gz> -C data

3 - Model checkpoints

Checkpoints for all trained models are available here.

Please untar them into the checkpoints folder as follows:

tar -xzvf <path-to-tar.gz> -C checkpoints

4 - Run!

Once your setup is complete, running tests is as simple as running test.py.

Usage:

usage: test.py [-h]

positional arguments:
              The name of the dataset to perform tests on.Choose among
              `mnist`, `cifar10`, `ucsd-ped2`, `shanghaitech`

optional arguments:
  -h, --help  show this help message and exit

Example:

python test.py ucsd-ped2