Self-Supervised Anomaly Segmentation

Intorduction

This is a PyToch implementation of A Novel Self-supervised Learning Task Designed for Anomaly Segmentation, it contains two mainly parts, Pseudo Mask Generator and Anomaly Segmentation Framework, as shown in next:

Pseudo Mask Generator:

Anomaly Segmentation Framework(ResNet50+FPN+DBNet):

Contributions

• we propose a novel self-supervised learning pretext task, which is different from generation-based methods or commonly contrastive leanring, it generat pseudo mask from other labeled dataset such as CoCo, and every suitable for pixelwise downstream tasks.
• we present an end-to-end anomaly segmenation framework, it has both high speed and accuracy, and with no post-processing.
• our method achieve SOTA in three anomaly detection/segmentation datasets. (#ToDo)

Anomaly Segmentation Demo(SHTech dataset)

Dataset Download

• USCD Ped2 [dataset]
• CUHK Avenue [dataset]
• ShanghaiTech [dataset]
• COCO [dataset]

Installation and Usage

1. prepare environment:

   conda create -n ssas python=3.7.6
   conda activate ssas
   pip install -r requirements.txt
   git clone https://github.com/wufan-tb/ssas
   

2. prepare coco pseudo mask:

   cd dataset
   python select_coco_annotation.py --image_dir {coco img folder} --annotation_path {coco_annotation.json}
   cd ..
   

3. training vad dataset(such as Ped2, SHTech):

   python train.py --dataset_path {your dataset path}
   

4. evaluation:

   python eval.py --dataset_path {your dataset path}
   

5. testing(generating segmentation demo):

   python inference.py --input {test imgs or video or camera} --output {save dir} --weights {xxx.pt}
   

Training Sample

Citation

If you find our work useful, please cite as follow:

{   ssas,
    author = {Wu Fan},
    title = { Self-Supervised Anomaly Segmentation },
    year = {2021},
    url = {\url{https://github.com/wufan-tb/ssas}}
}