Fast Dynamic Prototypes (FDP)

PyTorch implementation of Fast Dynamic Prototypes for Unsupervised Anomaly Detection and Localization (ECCV 2026).

1. Environment

conda create -n fdp python==3.10.16 -y
conda activate fdp
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu126
pip install -r requirements.txt

2. Dataset and Pre-trained weights

2.1 Dataset preparation

Convert each source dataset into its direct destination below data/:

python utils/mvtec3d2mvtecad.py --source <mvtec-3d-source> --target data/mvtec3d
python utils/visa2mvtecad.py --split_type 1cls --data_dir <visa-source> --save_dir data/visa --split_csv <visa-source>/split_csv/1cls.csv
python utils/realiad2mvtecad.py --json_dir <realiad-source>/realiad_jsons --source_dir <realiad-source>/realiad_1024 --output_dir data/realiad

The expected processed category layout is:

data/<dataset>/<category>/
|-- train/good/*.png
|-- test/good/*.png
|-- test/<defect-type>/*.png
`-- ground_truth/<defect-type>/*_mask.png

We provide pre-processed data available at: Download Link

2.2 Backbone weights

By default, DINOv3-small is loaded from encoder_weights/small.safetensors. If the local file is unavailable and network download is desired, set this value in Configs.py:

DEFAULT_ALLOW_BACKBONE_DOWNLOAD = True

This path requires timm>=1.0.20. The auxiliary ResNet18 backbone is loaded through torchvision and may download ImageNet weights on first use.

We provide pre-downloaded weights available at: Download Link

2.3 Foreground masks

Generate BiRefNet foreground masks into cache/<dataset>:

python utils/generate_foreground_masks.py \
  --data_dir data/mvtecad \
  --output_root cache/mvtecad \
  --target-category bottle cable capsule hazelnut metal_nut pill screw toothbrush transistor zipper \
  --device cuda:0

We provide pre-processed foreground mask available at: Download Link

2.4 Directory layout

FDP/
|-- data/
|   |-- dtd/images/...
|   |-- mvtecad/<category>/{train,test,ground_truth}/...
|   |-- mvtec3d/<category>/{train,test,ground_truth}/...
|   |-- visa/<category>/{train,test,ground_truth}/...
|   `-- realiad/<category>/{train,test,ground_truth}/...
|-- cache/
|   `-- <dataset>/<category>/train/good/...
|-- encoder_weights/
|   |-- small.safetensors
|   `-- modelscope/...
`-- logs/
    `-- <dataset>/
        |-- prototype_banks/
        |   `-- <category>_8192.pth
        `-- <mode>/
            |-- training/
            |-- checkpoints/
            |   |-- fdp/<category>/*.pth
            |   `-- seghead/<category>/*.pth
            |-- results/metrics.json
            `-- samples/<category>/<checkpoint>/*.png

3. Central configuration

All project paths are defined in Configs.py; no project path is read from an environment variable. Edit the user-settings block directly before running an experiment:

PROJECT_ROOT = Path(__file__).resolve().parent
DATA_ROOT = PROJECT_ROOT / 'data'
CACHE_ROOT = PROJECT_ROOT / 'cache'
LOGS_ROOT = PROJECT_ROOT / 'logs'
ENCODER_WEIGHTS_ROOT = PROJECT_ROOT / 'encoder_weights'

DEFAULT_DATASET = 'mvtecad'          # mvtecad / mvtec3d / visa / realiad
DEFAULT_MODE = 'multi'               # multi / single / cross
DEFAULT_DEVICE = 'cuda:0'
DEFAULT_BATCH_SIZE = 16
DEFAULT_NUM_WORKERS = 8
DEFAULT_ALLOW_BACKBONE_DOWNLOAD = False
DEFAULT_SAVE_VISUALIZATION_SAMPLES = True
DEFAULT_NUM_VISUALIZATION_SAMPLES = 8

4. Train and evaluate

Set DEFAULT_DATASET and DEFAULT_MODE in Configs.py, then run:

python prototypes_bank.py
python train.py
python train_seg.py
python test.py
python results.py

We provide trained weights available at: Download Link

Field	Default	Meaning
batch_size	2	Batch size
num_workers	8	Data-loader workers
num_visualization_samples	8	Test samples saved per class and checkpoint under logs/.../samples/