Wild SPADE : The Fast and the Furious 🏎🔥

This is an implementation of the paper Sub-Image Anomaly Detection with Deep Pyramid Correspondences.

We measured accuracy and speed for k=3, k=5 and k=50.

This code was implemented with reference to SPADE-pytorch, thanks.

Prerequisites

• faiss-gpu (easy to install with conda : ref)
• torch
• torchvision
• numpy
• opencv-python
• scipy
• argparse
• matplotlib
• scikit-learn
• torchinfo
• tqdm

Install prerequisites with:

conda install --file requirements.txt

Please download MVTec AD dataset.

After downloading, place the data as follows:

./
├── main.py
└── mvtec_anomaly_detection
    ├── bottle
    ├── cable
    ├── capsule
    ├── carpet
    ├── grid
    ├── hazelnut
    ├── leather
    ├── metal_nut
    ├── pill
    ├── screw
    ├── tile
    ├── toothbrush
    ├── transistor
    ├── wood
    └── zipper

Usage

To test SPADE on MVTec AD dataset:

python main.py

After running the code above, you can see the ROCAUC results in result/roc_curve.png

Results

Below is the implementation result of the test set ROCAUC on the MVTec AD dataset.

1. Image-level anomaly detection accuracy (ROCAUC %)

2. Pixel-level anomaly detection accuracy (ROCAUC %)

3. Processing time (sec)

CPU : Intel Xeon Platinum 8360Y
GPU : NVIDIA A100 SXM4

ROC Curve

• k = 3

• k = 5

• k = 50

Prediction Distribution (k = 5)

• bottle

• cable

• capsule

• carpet

• grid

• hazelnut

• leather

• metal_nut

• pill

• screw

• tile

• toothbrush

• transistor

• wood

• zipper

Localization (k = 5)

• bottle (test case : broken_large)

• cable (test case : bent_wire)

• capsule (test case : crack)

• carpet (test case : color)

• grid (test case : bent)

• hazelnut (test case : crack)

• leather (test case : color)

• metal_nut (test case : bent)

• pill (test case : color)

• screw (test case : manipulated_front)

• tile (test case : crack)

• toothbrush (test case : defective)

• transistor (test case : bent_lead)

• wood (test case : color)

• zipper (test case : broken_teeth)

For your infomation

We also implement a similar algorithm, PatchCore.
https://github.com/any-tech/PatchCore-ex/tree/main