D2Rec

Official PyTorch Implementation of Dual-Masked and Discriminative Reconstruction for Unified Vision Anomaly Detection, IEEE TIP.

Updates

• Support training and evaluation for large-scale Real-IAD-Variety
• Add the results of MVTec, VisA, BTAD, Medical and Real-IAD-Variety

Introduction

D2Rec is a simple, effective, general and robust unified (multi-class) vision anomaly detection framework that integrates unsupervised dual-masked reconstruction and a self-supervised discriminator, achieving competitive performance on both industrial and medical anomaly detection benchmarks.

D2Rec Framework

Main Results

Evaluation with 224x224 input resolution and the last checkpoint (the 50-th epoch).

datasets	#classes	#test images	I-AUROC	P-AUROC	I-AUPR	P-AUPR
MVTec	15	1725	98.9	99.6	98.9	74.3
VisA	12	2162	95.4	96.3	99.0	48.5
BTAD	3	741	96.2	96.6	97.6	61.3
Medical	3	7013	88.6	88.5	98.0	60.6
Real-IAD-Variety	160	178995	88.1	97.7	93.8	46.5

Evaluation with 448x448 input resolution and the last checkpoint (the 50-th epoch).

datasets	#classes	#test images	I-AUROC	P-AUROC	I-AUPR	P-AUPR
MVTec	15	1725	99.3	99.7	98.6	77.6
VisA	12	2162	97.2	97.9	98.6	53.6
BTAD	3	741	95.5	97.3	97.5	65.5
Medical	3	7013	89.0	88.9	97.3	62.0
Real-IAD-Variety	160	178995	84.5	97.0	92.9	45.5

Please see more detailed results in the results folder.

1. Environments

Create a new conda environment and install required packages.

conda create -n d2rec python=3.8.12
conda activate d2rec
pip install -r requirements.txt

2. Prepare Datasets

Download MVTec, VisA, BTAD, Medical and Real-IAD-Variety datasets from the official websites and unzip them to ./data/.

You can freely use the provided 'meta.json' files in './data'. You can also use the scripts in ./gen_meta_json/ to generate meta.json for each dataset with the following command:

python3 ./gen_meta_json/mvtec.py
python3 ./gen_meta_json/visa.py 
python3 ./gen_meta_json/btad.py
python3 ./gen_meta_json/medical.py 
python3 ./gen_meta_json/real-iad-variety.py

3. Training

using unified (i.e., multi-class) vision anomaly setting

image_size=224
for dataset in mvtec visa btad medical Real-IAD-Variety
do 
CUDA_VISIBLE_DEVICES=0 python3 main.py  \
   --data_path   "./datasets/"$dataset \
   --dataset $dataset \
   --image_size ${image_size} \
   --batch_size 16 \
   --dual_mask \
   --mask_head
done

4. Evaluation

image_size=224
for dataset in mvtec visa btad medical  Real-IAD-Variety
do 
CUDA_VISIBLE_DEVICES=0 python3 main.py  \
   -e \
   --data_path   "./datasets/"$dataset \
   --dataset $dataset \
   --save_path  "./checkpoints/" \
   --image_size ${image_size} \
   --batch_size 16 \
   --dual_mask \
   --mask_head
done

Citing

If you find this code useful in your research, please consider citing us:

@article{gao2026d2rec,
  title  = {Dual-Masked and Discriminative Reconstruction for Unified Vision Anomaly Detection},
  author = {Gao, Bin-Bin},
  booktitle = {IEEE Transactions on Image Processing},
  pages = {4701-4712},
  year = {2026}
}

Star History