RGB Ensemble Strategies for Unsupervised Industrial Anomaly Detection on the AutoVI Dataset

This repository contains the source code and experimental results for reproducing the experiments presented in our paper.

Authors: Sergio Villanueva, Emilio Soria-Olivas, Manuel Sanchez-Montanes

Overview

We evaluate nine unsupervised anomaly detection (UAD) models on the AutoVI dataset (automotive inspection) and propose late-fusion ensemble strategies to improve detection and localization performance. AutoVI is a publicly available dataset introduced by Carvalho et al. (2024) in Computers in Industry.

Models evaluated: PatchCore, PaDiM, EfficientAD, FastFlow, Reverse Distillation, DiNomaly, U-Flow, SuperSimpleNet, UniNet

Ensemble strategies: Mean, Median, Max, Trimmed Mean, Rank Average (applied at both score and pixel-map levels)

Requirements

• Python 3.11 (tested with 3.11.x)
• NVIDIA GPU with CUDA 12.x (tested on RTX A5000, 24GB VRAM)

Install dependencies (recommended: uv):

uv sync

If you prefer plain pip, install the same dependencies listed in pyproject.toml.

Dataset

Download the AutoVI dataset (Carvalho et al., Computers in Industry, 2024) and place it under data/ following the MVTec-style structure:

data/
├── engine_wiring/
│   ├── train/good/
│   ├── test/good/
│   ├── test/<defect_type>/
│   └── ground_truth/<defect_type>/
├── pipe_clip/
├── pipe_staple/
├── tank_screw/
├── underbody_pipes/
└── underbody_screw/

For cross-modal experiments, generate pseudo-depth maps using Depth Anything V2 and place them in data/<category>_estimate_depth/ with the same structure.

Reproducing Experiments

1. Train individual models

uv run train_all.py --device 0

This trains all models on all 12 datasets (6 RGB + 6 pseudo-depth) with 3 random seeds (1, 2, 42). Results are saved to results_0/ following this structure:

results_0/all_models/<dataset>/<seed>/<model>/...

The paper focuses on a subset of comparisons in some analyses, but the code supports all models listed above.

Training takes approximately 48-72 hours on a single RTX A5000.

2. Run within-modality ensembles (RGB)

uv run ensemble_simple.py

Combines model predictions using late-fusion strategies. Results are saved to ensemble_results/.

3. Run cross-modal ensembles (RGB + Depth)

uv run cross_modal_ensemble.py

Evaluates RGB + pseudo-depth fusion strategies. Results are saved to cross_modal_results/.

4. Analyze results

uv run ensemble_analysis.py
uv run data_analysis.py

Generates summary tables and statistical analysis.

5. Generate paper tables and figures

(Note: If you have skipped training and ensemble generation due to dataset size, you can directly generate the final tables and figures using the pre-computed CSVs in results/)

uv run tables_to_latex.py
uv run csv_to_tables.py
uv run graph_ensemble.py

Pre-computed Results

The results/ directory contains the pre-computed metrics from our experiments (no images are redistributed):

If you only want to regenerate the paper tables/figures from these CSVs (without training and without downloading AutoVI), a lightweight setup is enough:

pip install numpy pandas scipy matplotlib
python data_analysis.py
python csv_to_tables.py
python tables_to_latex.py

File	Description
training_results.csv	Raw per-seed results for all individual models (semicolon-delimited, comma decimal)
paper_table2_summary_with_std.csv	Aggregated mean and std across seeds
paper_table1_significant_improvements.csv	Ensemble improvements over best individual models
paper_table3_best_per_dataset.csv	Best ensemble configuration per dataset
complete_ensemble_analysis.csv	Full seed-level ensemble analysis

Project Structure

├── train_all.py              # Main training script (individual UAD models)
├── custom_metrics.py          # sPRO, AUsPRO, TPR@TNR metric implementations
├── ensemble_simple.py         # Within-modality late-fusion ensembles
├── cross_modal_ensemble.py    # Cross-modal RGB+Depth fusion
├── ensemble_analysis.py       # Post-hoc analysis of ensemble results
├── data_analysis.py           # Statistical analysis and figure data
├── tables_to_latex.py         # LaTeX table generation
├── csv_to_tables.py           # Pareto frontier plots
├── graph_ensemble.py          # Ensemble comparison visualizations
├── export_overlays.py         # Anomaly heatmap overlay generation
├── mosaics.py                 # Multi-panel figure generation
├── pyproject.toml             # Python dependencies and project metadata
└── results/                   # Pre-computed experimental results

Key Configuration

Parameter	Value
Framework	Anomalib 2.1
Seeds	{1, 2, 42}
Max epochs	100
Early stopping	patience=12, min_delta=1e-5
Precision	FP16 (mixed) for most models
PatchCore train limit	128 images

Metrics

Image-level: AUROC, AP (AUPR), TPR@TNR (at 99%, 95%, 90%)

Pixel-level: sPRO@0.05, AUsPRO@0.05 (Saturated Per-Region Overlap at 5% FPR)

Citation

If you use this code, please cite:

@article{villanueva2025rgb,
  title={RGB Ensemble Strategies for Unsupervised Industrial Anomaly Detection on the {AutoVI} Dataset},
  author={Villanueva, Sergio and Soria-Olivas, Emilio and Sanchez-Montanes, Manuel},
  journal={Computers in Industry},
  year={2025}
}

License

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