ProCon: Training-Free Anomaly Detection via Depth-Selective Soft-Projection Consensus

ProCon (Projection-Consensus, a.k.a. LayerConsensus) is a training-free unsupervised anomaly detection (UAD) method. It improves retrieval-based UAD purely by redesigning the memory bank and the scoring rule — no decoder training, no backbone fine-tuning, no pseudo-anomaly supervision — on top of a frozen DINOv2 ViT-B/14.

Each of a small pool of transformer layers {4, 5, 7, 10} (1-based, of 12) keeps its own independent 1% coreset memory and produces a soft-projection reconstruction-residual map; the per-layer maps are fused by a fixed mean. This double consensus (bank-consensus nested inside layer-consensus) beats the soft-projection baseline on every pixel metric at the same 1% budget.

Highlights

• Training-free: the only "training" is greedy k-center coreset selection on normal features.
• Frozen backbone: a single forward pass of DINOv2 ViT-B/14; nothing is fine-tuned.
• 1% memory budget, identical to PatchCore.
• Generalizes across 6 benchmarks with the unchanged recipe (see table below).

Method

PatchCore        hard NN retrieval, single memory bank
  + bank axis  → bank consensus (median over B seed-perturbed coresets)
  + soft proj  → soft-projection residual   r = || z − Σ_j w_j m_j ||,  w = softmax(−d²/τ)
  + layer axis → ProCon: run the residual per layer on independent memory, mean-fuse the maps

Each layer produces its own residual map; averaging the depth-separated maps blends the image-level signal (deep layers) with localization (mid layers).

Full derivation, ablations, and per-category tables: docs/METHOD.md.

Results

All numbers are category-averaged, seed 0, with the same recipe on every dataset. The default operating point is a 1% coreset (identical budget to PatchCore).

dataset	#cat	coreset	I-AUROC	P-AUROC	P-AP	AUPRO
MVTec-AD	15	1%	0.9971	0.9862	0.7298	0.9566
VisA	12	1%	0.9910	0.9903	0.5229	0.9695
Real-IAD (single-view)	30	1%	0.9315	0.9904	0.4935	0.9719
MPDD	6	1%	0.9740	0.9786	0.5277	0.9359
BTAD	3	1%	0.9515	0.9778	0.7137	0.9292
Uni-Medical (BMAD, pixel)	3	1%	0.8767	0.9716	0.5594	0.9075

Coreset budget (MVTec-AD / VisA, all 8 metrics)

ProCon improves with the coreset budget and peaks at 5–10%, yet 1% already beats the soft-projection baseline at 10% — memory-bank design matters more than budget. Best value per column in bold.

dataset	coreset	I-AUROC	I-AP	I-F1	P-AUROC	P-AP	P-F1	AUPRO	PRO
MVTec-AD	1%	0.9971	0.9990	0.9924	0.9862	0.7298	0.7056	0.9566	0.9274
MVTec-AD	5%	0.9975	0.9992	0.9932	0.9869	0.7347	0.7092	0.9586	0.9338
MVTec-AD	10%	0.9976	0.9993	0.9940	0.9870	0.7355	0.7099	0.9588	0.9273
VisA	1%	0.9910	0.9924	0.9713	0.9903	0.5229	0.5493	0.9695	0.9030
VisA	5%	0.9919	0.9930	0.9746	0.9907	0.5228	0.5472	0.9703	0.8959
VisA	10%	0.9915	0.9927	0.9742	0.9908	0.5232	0.5468	0.9704	0.8950

Full 8-metric and per-category breakdowns (all datasets and budgets) are in docs/METHOD.md.

Qualitative

Input · ground truth · nearest-neighbor memory · soft-projection memory · ProCon:

Installation

conda create -n ad_env python=3.10 -y
conda activate ad_env
pip install -U pip
pip install -r requirements.txt
pip install -e .          # installs the `skipcore` package

Tested with Python 3.10, torch 2.5.1 + CUDA 12.1, torchvision 0.20.1 on a single 24 GB GPU.

Datasets

MVTec/VisA-style layout is expected (datasets are not included in this repo):

<root>/<category>/train/good/*.png
<root>/<category>/test/<defect_type>/*.png
<root>/<category>/ground_truth/<defect_type>/*_mask.png   # optional

Dataset roots are set per benchmark in configs/*.yaml. Supported: mvtec, visa, realiad, mpdd, btad, uni_medical.

Reproduce

# MVTec-AD + VisA (champion recipe, full 8-metric report)
bash scripts/reproduce_champion.sh

# Cross-domain benchmarks (MPDD, BTAD, Uni-Medical)
bash scripts/run_extra_benchmarks.sh

# Real-IAD, all 30 categories (single-view)
bash scripts/realiad_champion.sh

Or a single dataset directly:

python run_consensuscore.py --dataset mvtec --recipe p3_drop4_3689 --output runs/mvtec

Repository layout

run_consensuscore.py   # main entry point (build coreset + evaluate)
skipcore/              # core package
  consensus/           #   soft-projection scoring, layer-consensus runner, recipes
  models/backbones/    #   frozen DINOv2 multi-layer extractor
  memory/              #   approximate greedy k-center coreset
  data/ eval/ inference/ postprocess/ utils/
configs/               # per-dataset YAML configs
scripts/               # reproduction scripts
tools/                 # figure rendering + verification utilities
figures/               # figures
docs/METHOD.md         # full method + all benchmark results