MetaUAS

Unofficial PyTorch implementation of MetaUAS: Universal Anomaly Segmentation with One-Prompt Meta-Learning.

🖼️ Framework

🚀 Usage

1. 🔧 COCO Data Preprocessing

Install LaMa and download the pretrained model:

git clone https://github.com/advimman/lama.git
pip install -r lama/requirements.txt

curl -LJO https://huggingface.co/smartywu/big-lama/resolve/main/big-lama.zip
unzip big-lama.zip -d lama/big-lama/

Then prepare COCO train2017 images and annotations:

# Expected structure:
# /path/to/datasets/images/train2017/   # COCO train images
# /path/to/datasets/images/annotations/ # instances_train2017.json

# Generate CYWS coco-inpainted dataset
bash scripts/run_generate_cyws_dataset.sh

Adjust paths in scripts/run_generate_cyws_dataset.sh before running.

Alternatively, you can directly download the pre-generated COCO inpainted dataset from CYWS.

2. 🔥 Training

pip install -r requirements.txt
bash scripts/train_on_4gpu.sh

3. 📈 Evaluation

Two evaluation modes are supported:

• Oneprompt: one random normal sample per class as prompt, multi-seed averaging
• TopK: per-image top-k most similar normal samples as prompts

# Download pretrained checkpoint
wget https://huggingface.co/ldl010302/MetaUAS/resolve/main/metauas-256.pth

# Run both modes on MVTec AD and VisA
bash scripts/run_eval_mvtec.sh

Adjust CKPT, MVTEC_ROOT, VISA_ROOT in scripts/run_eval_mvtec.sh before running.

📊 Performance & Checkpoints

Experiments run on 4× NVIDIA RTX 3090, batch size 96/GPU, lr=1e-4, weight decay=0.005.

Results

Methods	Categories	Anomaly Classification			Anomaly Segmentation
		I-ROC	I-PR	I-F1max	P-ROC	P-PR	P-F1max	P-PRO
MetaUAS	bottle	98.9±0.9	99.7±0.2	98.3±0.9	98.3±0.7	84.4±2.0	75.9±1.2	93.6±1.4
	cable	91.1±1.9	95.2±0.9	86.9±1.4	94.3±0.3	66.7±1.0	63.4±1.2	86.7±1.3
	capsule	69.9±3.7	90.6±3.5	92.1±1.4	93.8±0.6	28.4±8.9	34.5±5.5	62.4±2.7
	carpet	99.6±0.2	99.9±0.0	99.0±0.4	98.1±0.2	68.8±1.0	64.3±0.5	96.0±0.6
	grid	91.7±1.2	97.3±0.4	90.2±1.3	93.0±1.7	26.8±3.6	33.4±3.2	74.7±5.8
	hazelnut	78.8±11.9	84.8±11.2	84.5±2.6	96.0±1.3	34.0±13.1	39.7±9.7	80.6±5.8
	leather	100.0±0.0	100.0±0.0	100.0±0.0	99.4±0.1	61.3±1.9	56.4±1.4	99.0±0.1
	metal nut	95.5±2.0	98.9±0.5	95.4±1.4	96.4±0.7	80.4±3.7	73.7±2.8	90.5±2.2
	pill	90.2±2.8	98.1±0.6	93.2±0.9	95.6±0.9	64.3±5.2	59.8±3.5	92.5±0.7
	screw	53.6±5.2	76.7±3.5	85.5±0.4	92.0±1.6	4.6±1.2	9.6±2.1	71.2±4.4
	tile	96.0±1.3	98.7±0.4	94.6±1.6	94.0±1.4	75.8±1.9	69.6±1.3	87.0±2.7
	toothbrush	92.8±1.6	97.4±0.6	92.0±1.7	98.4±0.3	58.8±2.4	59.6±2.9	83.8±1.2
	transistor	83.9±5.4	82.4±4.6	74.6±5.9	85.7±3.4	40.6±5.3	40.7±5.4	75.2±5.2
	wood	99.8±0.2	99.9±0.1	99.2±0.0	94.1±0.7	68.7±1.6	64.7±1.2	93.9±0.6
	zipper	95.0±2.8	98.5±0.9	95.3±2.0	97.1±1.1	58.9±4.5	55.7±3.1	69.7±5.2
	mean	89.1±1.1	94.5±1.0	92.1±0.6	95.1±0.2	54.8±0.9	53.4±0.6	83.8±0.6

Methods	Categories	Anomaly Classification			Anomaly Segmentation
		I-ROC	I-PR	I-F1max	P-ROC	P-PR	P-F1max	P-PRO
MetaUAS*	bottle	99.5	99.9	98.4	98.4	83.9	75.2	92.9
	cable	96.3	98.0	93.4	96.4	70.5	66.1	91.2
	capsule	89.3	97.5	93.7	95.4	45.7	46.4	68.5
	carpet	99.7	99.9	98.9	97.9	68.2	64.0	95.4
	grid	95.7	98.7	93.7	94.0	29.4	36.8	79.7
	hazelnut	99.6	99.8	97.9	98.8	68.2	66.9	92.8
	leather	100.0	100.0	100.0	99.4	61.1	57.0	98.8
	metal nut	97.2	99.4	96.3	97.2	83.1	76.6	93.6
	pill	82.5	96.7	91.6	94.0	58.6	57.4	91.0
	screw	82.4	92.7	87.0	97.4	23.6	27.9	65.4
	tile	96.6	98.9	94.4	94.7	76.7	71.1	88.7
	toothbrush	96.4	98.6	93.8	99.0	64.3	63.2	81.2
	transistor	89.0	86.3	79.1	87.9	45.2	45.0	79.3
	wood	99.7	99.9	98.4	93.2	67.0	63.0	93.7
	zipper	93.6	98.1	95.0	96.8	58.4	56.2	68.2
	mean	94.5	97.6	94.1	96.0	60.2	58.2	85.4

Methods	Categories	Anomaly Classification			Anomaly Segmentation
		I-ROC	I-PR	I-F1max	P-ROC	P-PR	P-F1max	P-PRO
MetaUAS	candle	90.1±1.2	90.6±1.0	84.0±1.9	98.4±0.4	33.1±2.5	38.8±1.3	85.5±1.8
	capsules	62.7±5.5	73.4±4.0	78.1±0.6	92.2±2.2	16.0±2.8	24.7±1.9	52.2±2.3
	cashew	86.1±2.8	93.7±1.3	85.6±2.1	97.2±1.3	71.7±3.3	67.9±2.6	79.2±2.0
	chewinggum	96.9±1.0	98.7±0.4	94.5±0.6	99.5±0.1	82.5±1.2	76.6±0.8	85.3±0.9
	fryum	78.9±2.7	89.2±2.9	82.0±0.9	86.6±1.7	19.9±3.5	29.0±2.9	48.2±5.8
	macaroni1	77.2±1.4	80.7±0.8	71.9±1.7	92.6±2.4	12.7±0.7	21.6±0.6	56.9±4.2
	macaroni2	58.9±7.3	56.9±6.8	68.8±1.7	90.1±1.7	0.9±0.5	3.9±1.9	65.0±7.2
	pcb1	64.9±21.6	70.3±13.2	75.2±9.1	98.1±0.5	63.1±4.1	60.5±4.0	68.8±13.2
	pcb2	69.1±3.1	67.8±2.0	69.6±2.3	96.2±0.8	14.8±2.5	25.8±3.8	75.7±3.3
	pcb3	62.4±6.7	63.7±8.0	69.0±1.4	96.8±0.3	26.1±3.6	31.2±3.3	61.9±5.2
	pcb4	95.5±1.4	95.3±1.2	90.2±2.1	97.3±1.0	34.9±3.8	43.0±3.9	78.8±2.4
	pipe_fryum	95.4±1.9	97.5±1.1	93.6±1.9	98.7±0.5	67.8±3.3	63.2±2.2	85.8±1.3
	mean	78.2±2.0	81.5±1.5	80.2±1.0	95.3±0.5	37.0±0.9	40.5±0.6	70.3±1.5

Methods	Categories	Anomaly Classification			Anomaly Segmentation
		I-ROC	I-PR	I-F1max	P-ROC	P-PR	P-F1max	P-PRO
MetaUAS*	candle	91.2	91.2	84.7	98.6	35.2	39.1	84.9
	capsules	62.9	77.3	76.9	94.5	30.7	35.6	58.1
	cashew	85.5	93.4	84.3	98.6	77.9	71.4	76.8
	chewinggum	97.3	98.8	93.7	99.5	82.2	76.4	86.6
	fryum	79.5	89.9	81.7	88.9	22.3	31.7	37.7
	macaroni1	75.7	77.5	72.0	93.6	11.6	20.0	49.5
	macaroni2	59.2	56.3	68.1	91.7	1.1	5.6	67.8
	pcb1	82.5	80.5	77.1	99.2	76.3	71.3	69.3
	pcb2	70.0	69.6	67.7	97.2	14.8	26.3	79.0
	pcb3	76.2	75.8	73.6	97.2	31.1	35.5	57.0
	pcb4	95.3	95.6	88.5	96.8	35.6	43.8	72.2
	pipe_fryum	95.7	97.7	94.2	98.4	64.1	60.2	88.6
	mean	80.9	83.6	80.2	96.2	40.2	43.1	69.0

Pretrained Models

MetaUAS: metauas-256.pth

🙏 Acknowledgements

We reference code from MetaUAS, LaMa, and CYWS.

🤝 Contributing

• ⭐ If you find this project useful, a star would be greatly appreciated
• 🐛 Report bugs or ask questions via Issues
• 🔀 Fork and submit a PR with your improvements — we'll review and add you to the contributors