This is the official implementation of our paper, A Comprehensive Real-World Photometric Stereo Dataset for Unsupervised Anomaly Detection.
You can download the PSAD dataset through the following link.
The PSAD dataset is a new anomaly detection dataset with a photometric stereo set up. It contains over 10,000 high-resolution images divided into ten different object categories.
The normal map of our data set was created using the Woodham method. We provide light_directions.txt for each category, so if you want to create a normal map in a different way, you can use light_directions.txt. The light_directions.txt file is located in the light_calibration folder for each category. Note that we do not remove the background using a binary mask.
PSAD
│ README.md <- current readme page
│
├─images
│ ...
│
├─PSAD_DifferNet
│ │ ...
│ │ README.md <- explain how to run PSAD_DifferNet
│ │ ...
│ │
│
├─PSAD_MKDAD
│ │ ...
│ │ README.md <- explain how to run PSAD_MKDAD
│ │ ...
│
└─PSAD_skip_ganomaly
│ ...
│ README.md <- explain how to run PSAD_skip_ganomaly
│ ...
Benchmark models exist for each folder. (e.g. inside the PSAD_MKDAD folder for MKDAD models) Experiments can be reproduced for each model. Details are written in the README of each folder.
For example, if you want to reproduce the experimental results of the MKDAD model in our paper, click the PSAD_MKDAD folder and read README.md in the folder.
Note that the libraries of requirements (e.g. pytorch, torchvision) version is slightly different for each model.
If you find this useful for your research, please cite our paper:
@ARTICLE{9916256,
author={Jung, Junyong and Han, Seungoh and Park, Jinsun and Cho, Donghyeon},
journal={IEEE Access},
title={A Comprehensive Real-World Photometric Stereo Dataset for Unsupervised Anomaly Detection},
year={2022},
volume={10},
number={},
pages={108914-108923},
doi={10.1109/ACCESS.2022.3214003}}