Patch Purification for Contaminated Few-Shot Memory Banks

This repository contains the code and pre-computed results for the paper:

Patch purification for contaminated few-shot memory banks Sergio Villanueva, Emilio Soria-Olivas, Manuel Sanchez-Montañés

Memory-bank methods for visual anomaly detection store patch descriptors from a few normal reference images and score test patches by nearest-neighbor distance. In few-shot settings (5–20 references), even a single defective support image can inject anomalous patches into the bank and degrade performance. We address this with a training-free, patch-level self-purification step that removes suspicious patches before test-time scoring using leave-one-out consistency, Mahalanobis distance, or their ensemble.

Requirements

• Python 3.11
• CUDA-compatible GPU (for feature extraction only)
• ~45 GB disk space for datasets
• ~88 GB disk space for cached features (all 35 categories)

Install dependencies:

pip install -r requirements.txt

Install PyTorch (+ torchvision) separately depending on your hardware (CPU/CUDA):

• https://pytorch.org/get-started/locally/

Optional (speed): if you have FAISS available, experiments run faster. The code falls back to scikit-learn if FAISS is not installed.

Datasets

Download and extract the following datasets into a data/ directory:

Dataset	Categories	Download
MVTec AD	15	Link
VisA	12	Link
BTAD	3	Link
MVTec LOCO AD	5	Link

Expected structure:

data/
├── mvtec_AD/
│   ├── bottle/
│   │   ├── train/good/
│   │   └── test/{good,broken_large,...}/
│   ├── cable/
│   └── ...
├── VisA/
│   ├── candle/
│   ├── split_csv/1cls.csv
│   └── ...
├── btad/
│   ├── 01/
│   └── ...
└── mvtec_loco_AD/
    ├── breakfast_box/
    └── ...

Reproduction

Quick start (no datasets required)

You can reproduce the paper plots/tables directly from the pre-computed CSVs in output/:

python analysis_statistical_tests.py
python analysis_figures.py

Step 1: Feature extraction (GPU required)

Extract DINOv3 and/or CLIP features for all datasets:

python precache_features.py                  # DINOv3 ViT-L (default)
python precache_features.py --backbone clip  # CLIP ViT-L/14

Features are saved to output/feature_cache/ (DINOv3) or output/feature_cache_clip/ (CLIP). This step takes ~2 hours per backbone on a single GPU.

Step 2: Run experiments (CPU only)

# Main experiments: 35 categories, 5 seeds, contamination rates 10-30%
python run_experiments.py

# With CLIP backbone
python run_experiments.py --backbone clip

# LOF baseline (appends to main results)
python run_lof_baseline.py

# Ablation studies
python run_experiments.py --ablation percentile
python run_experiments.py --ablation knn_k

# Quick debug run (2 categories, 1 seed)
python run_experiments.py --debug

Results are saved as CSV files in output/exp_p3_002_full/ (DINOv3) or output/exp_p3_002_full_clip/ (CLIP).

Step 3: Generate figures and statistical tests

python analysis_figures.py
python analysis_statistical_tests.py
python analysis_heatmaps.py  # requires feature cache + ground-truth masks

Output figures and tables are saved to output/analysis/.

Pre-computed results

The output/ directory contains all pre-computed CSV results used in the paper. These are the same files that would be generated by running the experiments above, so the analysis scripts can read them directly without re-running experiments.

DINOv3 backbone (output/exp_p3_002_full/):

File	Description
results_v2.csv	Main results: 35 categories, 5 seeds, TL={10,20}, c={0,0.1,0.2,0.3}, all methods
results_extra_cont.csv	Extended contamination rates (c up to 50%)
results_tl5.csv	N=5 support set results
results_pixel_auroc.csv	Pixel-level AUROC and AUPR metrics
results_leakage_check.csv	Data integrity verification (Pool A/B split)
ablation_full_percentile.csv	Percentile threshold ablation (35 categories)
ablation_percentile.csv	Percentile threshold ablation (6 categories)
ablation_knn_k.csv	KNN k ablation

CLIP backbone (output/exp_p3_002_full_clip/):

File	Description
results_v2.csv	CLIP backbone results (same structure as DINOv3)

Key results

At contamination rate c=0.3 with N=10 support images (DINOv3 backbone):

Method	Mean AUROC	Delta vs Clean	Recovery
No purification	81.1%	-8.9	--
LOO consistency	84.7%	-5.3	40%
Mahalanobis	84.0%	-6.0	32%
Ensemble (LOO+Mahal)	85.7%	-4.3	51%
Oracle (clean bank)	90.0%	0.0	100%

Citation

@article{villanueva2025patchpurification,
  title={Patch purification for contaminated few-shot memory banks},
  author={Villanueva, Sergio and Soria-Olivas, Emilio and Sanchez-Monta{\~n}{\'e}s, Manuel},
  year={2025}
}

License

This project is licensed under the MIT License. See LICENSE for details.