GMM-DAE

A Pytorch Re-Implement Trial (Not Official) of paper: Video Anomaly Detection by Estimating Likelihood of Representations

Unfortunately, I haven't got the accuracy as good as the paper mentioned.（maybe I have missed some details, any help will be appreciated）

Here are something you may think it's valuable.

• a better object-detector result can get better AUROC, (for example, large image size, regard smaller object box, use different conf-thres(train > test))
• I get same result as the author in Ped2, yet big margin in Avenue and ShanghaiTech (nearly less ten point)
• Score Calculation and Score Smoothing can get a great improvement(you can find more details in this repo )

Requirements

pytorch >=1.5.0 ( I use 1.5.0 )

scikit-learn

Framework Overview

The framework include Three Parts:

1. dataset prepare, contain object detect(which I use yolov5) and computer dynamic image;
2. train denoised auto-encoder;
3. get feature cluster center(train GMM);
4. caculate anomaly score(evaluate);

Datasets

You can get the download link from here

Training:

1. prepare data

python prepare_dataset.py

2. train DAE

python train_DAE.py

3. train GMM

python train_GMM.py

details about parameters seen in scripts

Testing:

1. python evaluate.py 

   more visualization tools can find in notebook: test.ipynb

To Do List:

• try knowledge distillation
• try cluster loss

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

    {   GMM_DAE,
        author = {Wu Fan},
        title = { A Implementation of {GMM-DAE} Using {Pytorch}},
        year = {2020},
        howpublished = {\url{https://github.com/wufan-tb/gmm_dae}}
    }