Skip to content

rvorias/ind_knn_ad

Folders and files

NameName
Last commit message
Last commit date

Latest commit

Β 

History

67 Commits
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 

Repository files navigation

Industrial KNN-based Anomaly Detection

This repo aims to reproduce the results of the following KNN-based anomaly detection methods:

  1. SPADE (Cohen et al. 2021) - knn in z-space and distance to feature maps spade schematic
  2. PaDiM* (Defard et al. 2020) - distance to multivariate Gaussian of feature maps padim schematic
  3. PatchCore (Roth et al. 2021) - knn distance to avgpooled feature maps patchcore schematic

* actually does not have any knn mechanism, but shares many things implementation-wise.


Install

$ pipenv install -r requirements.txt

Usage

CLI:

$ python run.py METHOD [--dataset DATASET]

Results can be found under ./results/.

Code example:

from indad.model import SPADE

model = SPADE(k=5, backbone_name="resnet18")

# get some training data
class_name = "bottle"
train_ds, test_ds = MVTecDataset(class_name).get_dataloaders()

model.fit(train_ds)

# evaluate
image_rocauc, pixel_rocauc = model.evaluate(test_ds)
print(image_rocauc, pixel_rocauc)

Custom datasets

πŸ‘οΈ

Check out one of the downloaded MVTec datasets. Naming of images should correspond among folders. Right now there is no support for no ground truth pixel masks.

πŸ“‚datasets
 β”— πŸ“‚your_custom_dataset
  ┣ πŸ“‚ ground_truth/defective
  ┃ ┣ πŸ“‚ defect_type_1
  ┃ β”— πŸ“‚ defect_type_2
  ┣ πŸ“‚ test
  ┃ ┣ πŸ“‚ defect_type_1
  ┃ ┣ πŸ“‚ defect_type_2
  ┃ β”— πŸ“‚ good
  β”— πŸ“‚ train/good
$ python run.py METHOD --dataset your_custom_dataset

Results

πŸ“ = paper, πŸ‘‡ = this repo

Image-level

class SPADE πŸ“ SPADE πŸ‘‡ PaDiM πŸ“ PaDiM πŸ‘‡ PatchCore πŸ“ PatchCore πŸ‘‡
bottle - 98.8 99.8 β– 100.0β–  β– 100.0β– 
cable - 76.5 93.3 β– 99.5β–  96.2
capsule - 84.6 88.3 98.1 95.3
carpet - 84.3 β– 99.4 98.7 98.7
grid - 37.1 98.2 β– 98.2β–  93.0
hazelnut - 88.7 83.7 β– 100.0β–  100.0
leather - 97.1 99.9 β– 100.0β–  100.0
metal_nut - 74.6 99.4 β– 100.0β–  98.3
pill - 72.6 89.0 β– 96.6β–  92.8
screw - 53.1 83.0 98.1 96.7
tile - 97.8 98.6 98.7 β– 99.0β– 
toothbrush - 89.4 97.2 β– 100.0β–  98.1
transistor - 89.2 96.8 β– 100.0β–  99.7
wood - 98.3 98.9 β– 99.2β–  98.8
zipper - 96.7 89.5 β– 99.4β–  98.4
averages 85.5 82.6 95.3* 94.3 β– 99.1β–  97.7
  • PaDiM average referencing PaDiM-WR50-Rd550

Pixel-level

class SPADE πŸ“ SPADE πŸ‘‡ PaDiM πŸ“ PaDiM πŸ‘‡ PatchCore πŸ“ PatchCore πŸ‘‡
bottle 97.5 97.7 98.3 97.8 β– 98.6β–  97.8
cable 93.7 94.3 96.7 96.1 β– 98.5β–  97.4
capsule 97.6 98.6 98.5 98.3 β– 98.9β–  98.3
carpet 87.4 99.0 99.1 98.6 β– 99.1β–  98.3
grid 88.5 96.1 97.3 97.2 β– 98.7β–  96.7
hazelnut 98.4 98.1 98.2 97.5 β– 98.7β–  98.1
leather 97.2 99.2 99.2 98.7 β– 99.3β–  98.4
metal_nut β– 99.0β–  96.1 97.2 96.5 98.4 96.2
pill β– 99.1β–  93.5 95.7 93.2 97.6 98.7
screw 98.1 98.9 98.5 97.8 β– 99.4β–  98.4
tile β– 96.5β–  93.3 94.1 94.8 95.9 94.0
toothbrush β– 98.9β–  β– 98.9β–  98.8 98.3 98.7 98.1
transistor β– 97.9β–  96.3 97.5 97.2 96.4 97.5
wood 94.1 94.4 94.7 93.6 β– 95.1β–  91.9
zipper 96.5 98.2 98.5 97.4 β– 98.9β–  97.6
averages 96.9 96.8 97.5 96.9 β– 98.1β–  97.2

PatchCore-10 was used.

Hyperparams

The following parameters were used to calculate the results. They more or less correspond to the parameters used in the papers.

spade:
  backbone: wide_resnet50_2
  k: 50
padim:
  backbone: wide_resnet50_2
  d_reduced: 250
  epsilon: 0.04
patchcore:
  backbone: wide_resnet50_2
  f_coreset: 0.1
  n_reweight: 3

Progress

  • Datasets
  • Code skeleton
  • Config files
  • CLI
  • Logging
  • SPADE
  • PADIM
  • PatchCore
  • Add custom dataset option
  • Add dataset progress bar
  • Add schematics
  • Unit tests

Design considerations

  • Data is processed in single images to avoid batch statistics interference.
  • I decided to implement greedy kcenter from scratch and there is room for improvement.
  • torch.nn.AdaptiveAvgPool2d for feature map resizing, torch.nn.functional.interpolate for score map resizing.
  • GPU is used for backbones and coreset selection. GPU coreset selection currently runs at:
    • 400-500 it/s @ float32 (RTX3080)
    • 1000+ it/s @ float16 (RTX3080)

Streamlit demo

Run $ streamlit run streamlit_app.py


Acknowledgements

References

SPADE:

@misc{cohen2021subimage,
      title={Sub-Image Anomaly Detection with Deep Pyramid Correspondences}, 
      author={Niv Cohen and Yedid Hoshen},
      year={2021},
      eprint={2005.02357},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

PaDiM:

@misc{defard2020padim,
      title={PaDiM: a Patch Distribution Modeling Framework for Anomaly Detection and Localization}, 
      author={Thomas Defard and Aleksandr Setkov and Angelique Loesch and Romaric Audigier},
      year={2020},
      eprint={2011.08785},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

PatchCore:

@misc{roth2021total,
      title={Towards Total Recall in Industrial Anomaly Detection}, 
      author={Karsten Roth and Latha Pemula and Joaquin Zepeda and Bernhard SchΓΆlkopf and Thomas Brox and Peter Gehler},
      year={2021},
      eprint={2106.08265},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

MVTec dataset:

@article{Bergmann2021,
 author = {Paul Bergmann and Kilian Batzner and Michael Fauser and David Sattlegger and Carsten Steger},
 title = {The MVTec Anomaly Detection Dataset: A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection},
 journal = {International Journal of Computer Vision},
 year = {2021},
 volume = {129},
 number = {4},
 pages = {1038-1059},
 doi = {10.1007/s11263-020-01400-4}
}
@inproceedings{Bergmann2019,
 author = {Paul Bergmann and Michael Fauser and David Sattlegger and Carsten Steger},
 title = {MVTec AD β€” A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection},
 booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
 year = {2019},
 pages = {9584-9592},
 doi = {10.1109/CVPR.2019.00982}
}