Mahalanobis PatchCore

Mahalanobis PatchCore is a compact PatchCore-style anomaly detector for visual inspection. It keeps the nearest-neighbor scoring structure of PatchCore and adds a streaming-compatible Mahalanobis feature normalization stage before the memory bank is built.

The repository contains the runtime implementation and the MVTec AD experiment configs used for the baseline and ablation runs. The default experiments run on CPU and expect the MVTec AD categories under mvtec_datasets/.

Layout

• mhpc/: runtime package and active plugin implementations.
• configs/mvtec/baselines/: canonical MVTec baseline configs.
• configs/mvtec/ablations/: MVTec ablation configs.
• scripts/run_all_experiments.sh: batch runner for the provided configs.
• tests/: focused runtime and numerical checks.

Setup

Create an environment with Python 3.11 and install the pinned dependencies:

pip install -r requirements.txt
pip install -r requirements-dev.txt

The runtime uses faiss-cpu and CPU Torch wheels by default. GPU execution is not required for the provided configs.

Dataset

Place MVTec AD in a directory named mvtec_datasets at the repository root:

mvtec_datasets/
  bottle/
  cable/
  ...
  zipper/

Each category is expected to use the standard MVTec AD train/test/ground-truth layout.

Running An Experiment

Run a single config with:

python run_mhpc.py --config configs/mvtec/baselines/mvtec_streaming_mh_patchcore.yaml

Run all provided MVTec configs with:

bash scripts/run_all_experiments.sh --all --continue

Results are written under results/mvtec/<experiment>/<timestamp>/.

Tests

Run the focused runtime suite with:

pytest -q