Industrial KNN-based Anomaly Detection

⭐Now has streamlit support!⭐ Run $ streamlit run streamlit_app.py

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

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

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

Install

$ pipenv install -r requirements.txt

Usage

CLI:

$ python indad/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")

# feed healthy dataset
model.fit(...)

# get predictions
img_lvl_anom_score, pxl_lvl_anom_score = model.predict(...)

Custom datasets

👁️

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.

📂datasets
 ┗ 📂your_custom_dataset
  ┣ 📂 ground_truth/defective
  ┃ ┣ 📂 defect_type_1
  ┃ ┗ 📂 defect_type_2
  ┣ 📂 test
  ┃ ┣ 📂 defect_type_1
  ┃ ┣ 📂 defect_type_2
  ┃ ┗ 📂 good
  ┗ 📂 train/good

$ python indad/run.py METHOD --dataset your_custom_dataset

Results

📝 = paper, 👇 = this repo

Image-level

class	SPADE 📝	SPADE 👇	PaDiM 📝	PaDiM 👇	PatchCore 📝	PatchCore 👇
bottle	-	98.8	98.3	99.8	■100.0■	■100.0■
cable	-	76.5	96.7	93.3	■99.5■	96.2
capsule	-	84.6	■98.5■	88.3	98.1	95.3
carpet	-	84.3	99.1	■99.4	98.7	98.7
grid	-	37.1	97.3	98.2	■98.2■	93.0
hazelnut	-	88.7	98.2	83.7	■100.0■	100.0
leather	-	97.1	99.2	99.9	■100.0■	100.0
metal_nut	-	74.6	97.2	99.4	■100.0■	98.3
pill	-	72.6	95.7	89.0	■96.6■	92.8
screw	-	53.1	■98.5■	83.0	98.1	96.7
tile	-	97.8	94.1	98.6	98.7	■99.0■
toothbrush	-	89.4	98.8	97.2	■100.0■	98.1
transistor	-	89.2	97.5	96.8	■100.0■	99.7
wood	-	98.3	94.7	98.9	■99.2■	98.8
zipper	-	96.7	98.5	89.5	■99.4■	98.4
averages	85.5	82.6	97.5	94.3	■99.1■	97.7

Pixel-level

class	SPADE 📝	SPADE 👇	PaDiM 📝	PaDiM 👇	PatchCore 📝	PatchCore 👇
bottle	97.5	97.7	94.8	97.8	■98.6■	97.8
cable	93.7	94.3	88.8	96.1	■98.5■	97.4
capsule	97.6	98.6	93.5	98.3	■98.9■	98.3
carpet	87.4	99.0	96.2	98.6	■99.1■	98.3
grid	88.5	96.1	94.6	97.2	■98.7■	96.7
hazelnut	98.4	98.1	92.6	97.5	■98.7■	98.1
leather	97.2	99.2	97.8	98.7	■99.3■	98.4
metal_nut	■99.0■	96.1	85.6	96.5	98.4	96.2
pill	■99.1■	93.5	92.7	93.2	97.6	98.7
screw	98.1	98.9	94.4	97.8	■99.4■	98.4
tile	■96.5■	93.3	86.0	94.8	95.9	94.0
toothbrush	■98.9■	■98.9■	93.1	98.3	98.7	98.1
transistor	■97.9■	96.3	84.5	97.2	96.4	97.5
wood	94.1	94.4	91.1	93.6	■95.1■	91.9
zipper	96.5	98.2	95.9	97.4	■98.9■	97.6
averages	96.9	96.8	92.1	96.9	■98.1■	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

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)

Acknowledgements

hcw-00 for tipping sklearn.random_projection.SparseRandomProjection.
h1day for adding a custom range to the streamlit app.

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 Schölkopf and Thomas Brox and Peter Gehler},
      year={2021},
      eprint={2106.08265},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

cyj95/ind_knn_ad