Sub-Image Anomaly Detection with Deep Pyramid Correspondences (SPADE) in PyTorch

PyTorch implementation of Sub-Image Anomaly Detection with Deep Pyramid Correspondences (SPADE).

SPADE presents an anomaly segmentation approach which does not require a training stage.
It is fast, robust and achieves SOTA on MVTec AD dataset.

We used K=5 nearest neighbors, which differs from the original paper K=50.

Prerequisites

python 3.6+
PyTorch 1.5+
sklearn, matplotlib

Install prerequisites with:

pip install -r requirements.txt

If you already download MVTec AD dataset, move a file to data/mvtec_anomaly_detection.tar.xz.
If you don't have a dataset file, it will be automatically downloaded during the code running.

Usage

To test SPADE on MVTec AD dataset:

cd src
python main.py

After running the code above, you can see the ROCAUC results in src/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 %)

	Paper	Implementation
bottle	-	97.2
cable	-	84.8
capsule	-	89.7
carpet	-	92.8
grid	-	47.3
hazelnut	-	88.1
leather	-	95.4
metal_nut	-	71.0
pill	-	80.1
screw	-	66.7
tile	-	96.5
toothbrush	-	88.9
transistor	-	90.3
wood	-	95.8
zipper	-	96.6
Average	85.5	85.4

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

	Paper	Implementation
bottle	98.4	97.0
cable	97.2	92.3
capsule	99.0	98.4
carpet	97.5	98.9
grid	93.7	98.3
hazelnut	99.1	98.5
leather	97.6	99.3
metal_nut	98.1	97.1
pill	96.5	95.0
screw	98.9	99.1
tile	87.4	92.8
toothbrush	97.9	98.8
transistor	94.1	86.6
wood	88.5	95.3
zipper	96.5	98.6
Average	96.5	96.4