Industrial Anomaly Detection with Localization

A Comparative Study of PatchCore and PaDiM under Clean and Shifted Domains

---

👥 Authors

• Ivan Necerini (s345147)
• Jacopo Rialti (s346357)
• Fabio Veroli (s336301)

📁 Dataset Links

The datasets are available at the following links:

Dataset	Link	Description
MVTec AD (Clean)	Download	Original MVTec AD dataset
MVTec-Shift (Shifted)	No Link	Generated in notebook 03

The original MVTec AD dataset needs to be downloaded and uploaded to your Google Drive before running the notebooks. Before uploading the entire dataset, remove the other classes expect for the selected one (e.g., hazelnut, carpet, zipper), used in our study. The MVTec-Shift dataset will be generated in the third notebook using the provided code.

---

Notebooks Execution Order

Before running the notebooks, make sure to have the datasets downloaded and available in your Google Drive.

To reproduce the results, execute notebooks in sequential order for full reproducibility, all the common source code used in the notebooks is available in the folder src/. Load the notebooks in Colab and run them in order. The notebooks will create the necessary folders and files and results will be saved on your Google Drive at the end of the experiment.

Step	Notebook	Purpose	Output
1	01_data_exploration.ipynb	Dataset analysis and visualization	Exploratory plots
2	02_data_preparation.ipynb	Train/Val/Test split (80/20, seed=42)	data/processed/clean_splits.json
3	03_domain_shift_generation.ipynb	Generate MVTec-Shift dataset	data/shifted/, shifted_splits.json
4	04_patchcore_clean.ipynb	Train PatchCore on clean domain	outputs/models/patchcore_*_clean.npy
5	05_padim_clean.ipynb	Train PaDiM on clean domain	outputs/models/padim_*_clean.pt
6	06_evaluation_clean.ipynb	Evaluate both on Test-clean	outputs/results/clean_results.json
7	07_evaluation_shifted_no_adaptation.ipynb	Evaluate on Test-shift (no adaptation)	shifted_no_adaptation_results.json
8	08_evaluation_shifted_threshold_only.ipynb	Threshold-only adaptation	shifted_threshold_only_results.json
9	09_full_shift_adaptation.ipynb	Full model retraining on shifted	Models + results for full adaptation
10	10_global_model_clean.ipynb	Model-Unified setting	patchcore_global_clean.npy
11	11_coreset_ratio_analysis_on_shift.ipynb	Coreset ablation (1%, 5%, 10%)	Coreset comparison results

---

🗂️ Project Structure

AD/
├── configs/
│   └── experiment_config.yaml     # Parameters and Hyperparameters (seed=42, coreset=0.05, etc.)
├── data/
│   ├── processed/
│      ├── clean_splits.json      # Train/Val/Test indices for clean domain
│      └── shifted_splits.json    # Splits for shifted domain
├── notebooks/                     # Jupyter notebooks (01-11)
├── src/
│   ├── data/                      # Dataset, transforms, splitter
│   ├── models/                    # PatchCore wrapper, PaDiM wrapper
│   ├── metrics/                   # Evaluator, PRO metric
│   └── utils/                     # Config loader, visualization
├── outputs/
│   ├── models/                    # Trained models (.npy, .pt)
│   ├── results/                   # Evaluation metrics (.json, .csv)
│   ├── thresholds/                # Calibrated thresholds per experiment
│   └── visualizations/            # ROC curves, confusion matrices, heatmaps
├── report/                        # Paper PDF file
└── doc/                   # Submission materials

⚙️ Key Hyperparameters

Parameter	Value	Location
Random seed	42	configs/experiment_config.yaml
Input resolution	224×224	All notebooks
Backbone	ResNet-50 (ImageNet, frozen)	PatchCore/PaDiM
Coreset ratio	0.05 (default)	PatchCore
PaDiM n_features	100	PaDiM

---

📊 Supplementary Materials

Models

Here you can find all the trained models during the experiments:

Path	Content
outputs/models/patchcore_*_clean.npy	PatchCore memory banks (clean)
outputs/models/padim_*_clean.pt	PaDiM Gaussian parameters (clean)
outputs/models/*_shift.npy/.pt	Fully adapted models
outputs/models/*_coreset_01/10.*	Coreset ablation models

Results

Path	Content
outputs/results/clean_results.json	Clean domain metrics
outputs/results/shifted_*_results.json	All shift scenarios
outputs/results/global_results.json	Unified model metrics
outputs/results/patchcore_coreset_comparison.csv	Coreset ablation summary

Thresholds

Path	Content
outputs/thresholds/patchcore_clean_thresholds.json	Clean thresholds
outputs/thresholds/shift_threshold_only_*.json	Recalibrated thresholds

Visualizations

Path	Content
outputs/visualizations/06_evaluation_clean/	ROC curves, confusion matrices (clean)
outputs/visualizations/shifted_no_adaptation/	No-adaptation results
outputs/visualizations/shifted_full_adaptation/	Full adaptation results
outputs/visualizations/global/	Unified model + shortcut analysis
outputs/visualizations/coreset/	Coreset trade-off plots

---