Necolizer / Anomaly-Detection-Using-KolektorSDD-Dataset

Reorganize KolektorSDD dataset as MVTecAD dataset's format. Report SOTA anomaly detection models in KolektorSDD.

Anomaly-Detection-Using-KolektorSDD-Dataset

Reorganize KolektorSDD dataset as MVTecAD dataset's format. Report SOTA anomaly detection models' results in KolektorSDD.

0. Table of Contents

• 1. Purpose
• 2. Usage
• 3. Illustrations of KolektorSDD Preprocessing
• 4. Experimental Results
• 5. Sample Visualizations
• 6. Change Log
• 7. License

1. Purpose

zh

这是本仓库作者在实习期间完成的代码，主要内容是拿KolektorSDD数据集去跑部分SOTA工业缺陷检测模型（主要取自MVTecAD的排行榜）。由于各种限制，这里只公布了KolektorSDD的预处理代码和复现出的结果。通过KolektorSDD的预处理代码，可以快速将KolektorSDD数据集的格式转换为MVTecAD数据集的格式，这样就可以直接套用SOTA开源代码或者Anomalib库，快速的进行训练和测试。完整复现应该不困难。

en

This repository contains code for preprocessing KolektorSDD dataset so that we could train/test some SOTA anomaly detection models in MVTecAD leaderboard. Due to various restrictions, I do NOT upload the modified training/testing code for those SOTA models. But I believe with the KolektorSDD dataset after preprocessing, you could reproduce the results in a very short time, just with slight modifications to SOTA codes/Anomalib using MVTecAD configurations. I also report the results I reproduce for comparisons.

2. Usage

1. Download KolektorSDD Dataset with fine annotations in https://www.vicos.si/resources/kolektorsdd/
   • Please cite according to their requirements
2. Unzip the KolektorSDD file
3. Git clone this repo
4. Create a new conda virtual environment with anomalib_env.yaml

conda env create -f anomalib_env.yaml

5. Modify the path in KolektorSDD_Preprocess.py

# Args that you need to change: 
# @ read_base : Path to the downloaded KolektorSDD dataset
# @ save_base : Path to the repository you wanna save
read_base = r'.\KolektorSDD'
save_base = r'.\KolektorSDD1'

6. Run the script

python KolektorSDD_Preprocess.py

7. Your final directory tree of reorganized KolektorSDD should look like this:

save_base
└── metal
    ├── ground_truth
    |    └── defect
    |         ├── 000_mask.png
    |         ├── 001_mask.png
    |         ├── ...
    ├── test
    |    ├── defect
    |    |    ├── 000.png
    |    |    ├── 001.png
    |    |    ├── ...
    |    └── good
    |         ├── 000.png
    |         ├── 001.png
    |         ├── ...
    └── train
         └── good
             ├── 000.png
             ├── 001.png
             ├── ...

8. Modify the code configurations and run your training and testing scripts
   • See Section 4 to get the open source code for SOTA models in my experiment
   • With the KolektorSDD dataset after preprocessing, you could reproduce the results in a very short time, just with slight modifications to SOTA codes/Anomalib using MVTecAD configurations.

3. Illustrations of KolektorSDD Preprocessing

zh

总体思路：处理成接近MVTecAD数据集的样式

步骤：

1. Jpg和Bmp转Png
2. Resize到统一的尺寸：500x1240
3. 划分训练和测试
   • 训练：295张正常
   • 测试：52张正常+52张异常

这里需要注意：

• 本仓库作者采取的手段是直接Resize到统一的尺寸，这可能会导致某些小缺陷的mask变形，如果有时间，可以换成crop的形式，把有缺陷的部分crop出一个正方形区域出来。
• 本仓库作者采取的划分方式是直接划分训练和测试，没有留验证集。同时，在代码中已经规定了测试时的正常和异常样本数量相等。如果需要。可以自行修改代码，取合理的划分。
• 因为random.shuffle没有固定种子，每次运行会得到具体样本不同的划分结果。

en

To take use of SOTA codes / Anomalib using MVTecAD configurations, we should reorganize KolektorSDD dataset in MVTecAD dataset's format.

Steps:

1. JPG/BMP to PNG
2. Resize to the same size (500x1240)
3. Train-Test Split
   • Train
     • Flawless samples for training: 295
   • Test
     • Flawless samples for testing: 52
     • Anomalies for testing: 52

ATTENTION:

• I resize all the images to the same size (500x1240), which might result in defect distortions.
• Because of the small number of samples, I do NOT reserve a validation set. I sampled the same number of flawless samples as anomalies for testing. This setting could be changed as you want.
• Seed for random.shuffle() is NOT fixed. So each run of this preprocessing script will result in different splitting results.

4. Experimental Results

SOTA models that chosen to train/test (Also as Acknowledgements):

• PatchCore (backbone: wide_resnet50_2)
  • Official Code Ver.
  • Anomalib Ver.
• FastFlow (backbone: resnet18 / cait_m48_448)
  • Anomalib Ver.
  • Unofficial Code Ver.
• CFA (backbone: wrn50_2)
  • Official Code Ver.
• Cflow-AD (backbone: wide_resnet50_2)
  • Official Code Ver.
  • Anomalib Ver.

Version	Methods	Backbone	Avg DET AUC (image ROCAUC)	Avg SEG AUC (pixel ROCAUC)	pixel PROAUC
Official Code	PatchCore	wide_resnet50_2	0.909	0.941	/
	CFA	wrn50_2	0.939	0.939	0.823
	Cflow-AD	wide_resnet50_2	0.801	0.891	0.497
Unofficial Code	FastFlow	cait_m48_448	0.955	0.960	/
Anomalib	PatchCore	wide_resnet50_2	0.863	0.840	/
	FastFlow	resnet18	0.807	0.883	/
		cait_m48_448	0.914	0.963	/
	Cflow-AD	wide_resnet50_2	0.850	0.847	/

5. Sample Visualizations

• PatchCore (wide_resnet50_2)

• FastFlow (cait_m48_448)

• Cflow-AD (wide_resnet50_2)

• CFA (wrn50_2)

6. Change Log

• [2022/12/29] Provide conda virtual environment dependency list in anomalib_env.yaml.
• [2022/12/28] Create repository and release preprocessing script.

7. License

MIT