patchcore

PatchCore_anomaly_detection

Input

Normal images

(Image from MVTec AD datasets https://www.mvtec.com/company/research/datasets/mvtec-ad/)

• Original image shape : (n, 3, 900, 900)
• Input shape : (n, 3, 224, 224)

Output

Anomaly maps

Left to right: input, ground truth, predicted head map, predicted mask, segmentation result

Usage with MVTec AD datasets

In order to get the feature vector of the normal product, it is necessary to prepare the file of the normal product.

By default, normal files are got from the train directory.
For the sample image, first download MVTec AD datasets and place bottle/train/good/*.png files to the train directory.

For the sample image, train with train directory and test with bottle_000.png,

$ python3 patchcore.py

For train and test directly from a particular category in MVTec AD datasets.

$ python3 patchcore.py -i bottle/test/broken_large --train_dir bottle/train/good --gt_dir bottle/test/ground_truth/broken_large

Automatically downloads the onnx and prototxt files on the first run. It is necessary to be connected to the Internet while downloading.

Usage with your datasets

By default, mask image required to calculate the optimal threshold. By specifying the threshold option, it is not necessary to prepare the mask image.

With the following command, learn using train folder and verify with test folder.

$ python3 patchcore.py --train_dir train --input test --threshold 0.5

Use the following command to perform only the test.

$ python3 patchcore.py --feat train.pkl --input test --threshold 0.5

Now you can give videos to train_dir and video option. If a video is given, the first 200 frames of the video will be used for training.

$ python3 patchcore.py --train_dir train.mp4 --video test.mp4 --threshold 0.5

Options

You can specify the directory of normal product files with the --train_dir option.

$ python3 patchcore.py --train_dir train

The feature vectors created from files in the train directory are saved to the pickle file.
From the second time, by specifying the pickle file by --feat option, it can omit the calculation of the feature vector of the normal product.
The name of the pickle file created is the same as the name of a normal product file directory.

$ python3 patchcore.py --feat train.pkl

The ground truth files are got from the gt_masks directory by default.
The name of the ground truth file corresponds to the file with __mask after the name of the input file.
You can specify the directory of ground truth files with the --gt_dir option.

$ python3 patchcore.py --gt_dir gt_masks

If you want to specify the input test image, put the image path after the --input option.
You can use --savepath option to change the name of the output file to save.

$ python3 patchcore.py --input IMAGE_PATH --savepath SAVE_IMAGE_PATH

By adding the --arch option, you can specify model type which is selected from "resnet18", "wide_resnet50_2".
(default is resnet18)

$ python3 patchcore.py --arch wide_resnet50_2

By adding the --aug option, you can process with augmentation.
(default is processing without augmentation)

$ python3 patchcore.py --aug

Reference

PatchCore_anomaly_detection

Framework

Pytorch

Model Format

ONNX opset=11

Netron

resnet18.onnx.prototxt

wide_resnet50_2.onnx.prototxt