PatchCore Customization Notes (Kvasir)

This project adapts PatchCore to run on a custom Kvasir-based anomaly-detection dataset arranged in an MVTec-like structure.

What Was Changed

1. Custom dataset preparation pipeline

Added prepare_dataset.py to build a PatchCore-compatible folder layout from:

• kvasir-dataset-v2 (healthy images)
• Kvasir-SEG (polyp images + masks)

Configured paths in prepare_dataset.py:

• KVASIR_V2_DIR = ../dataset/kvasir-dataset-v2
• KVASIR_SEG_DIR = ../dataset/Kvasir-SEG
• OUTPUT_DIR = ../dataset/kvasir_patchcore

Healthy classes used:

• normal-cecum
• normal-pylorus
• normal-z-line

Train/test split:

• TRAIN_SPLIT = 0.80

Generated output tree:

kvasir_patchcore/
	train/
		good/
	test/
		good/
		defect/
	ground_truth/
		defect/

Behavior:

• Healthy images from Kvasir-v2 are shuffled with a fixed seed (42) and split into train/good and test/good.
• Polyp images from Kvasir-SEG are copied to test/defect.
• Matching segmentation masks are copied to ground_truth/defect.

2. Kvasir dataset loader updates

Updated patchcore-inspection/src/patchcore/datasets/kvasir.py:

• Default class list changed from MVTec classes to:
  • _CLASSNAMES = ["kvasir_patchcore"]
• Dataset docstring text updated from MVTec wording to Kvasir wording.
• Added:
  • self.transform_mean = IMAGENET_MEAN
  • self.transform_std = IMAGENET_STD

Purpose:

• Ensure the loader defaults to the generated Kvasir class folder.
• Keep normalization statistics available as dataset attributes.

Expected Dataset Location

After running preparation, the training/evaluation scripts should point to:

• ../dataset/kvasir_patchcore

Inside that path, PatchCore expects one class folder:

• kvasir_patchcore/
  • containing train, test, and ground_truth as shown above.

Quick Run Order

1. Prepare data:

python prepare_dataset.py

2. Train/evaluate PatchCore using the Kvasir dataset path:

• dataset root: ../dataset/kvasir_patchcore
• class name: kvasir_patchcore

Notes

• This setup treats Kvasir as a one-class anomaly-detection benchmark in PatchCore format.
• good = healthy endoscopy images; defect = polyp samples with masks.