AnomaVision
Simple visual anomaly detection from normal images.
AnomaVision is a production-oriented computer vision toolkit for detecting defects and unusual patterns from normal images.
Supported methods:
- PaDiM
- PatchCore
- EfficientAD
It supports training, evaluation, threshold calibration, ONNX/OpenVINO/TensorRT export where supported, KV260/XModel deployment, production data-drift monitoring, and non-invasive deployment validation.
🚀 Quick start
📦 Install
git clone https://github.com/DeepKnowledge1/AnomaVision.git
cd AnomaVision
uv venv --python 3.11 .venv
source .venv/bin/activate # Windows: .venv\\Scripts\\Activate.ps1
uv sync --extra cpu
🧠 Train
anomavision train --config config.yml
🔍 Detect
anomavision detect --config config.yml --img_path ./dataset/bottle/test
📤 Export
anomavision export --config config.yml --format onnx
🛡️ Deployment validation
Validate exported models before deployment without changing the anomaly-detection algorithm.
It checks model integrity, inference, performance, backend compatibility, and output consistency against a reference model.
anomavision validate `
--model model.onnx `
--reference-model model.pt `
--config config.yml
🛡️ Deployment validation
Before moving an exported model into production, AnomaVision can validate the deployment artifact without changing the underlying anomaly-detection algorithm.
It checks model integrity, inference, performance, supported backend availability, and—when a reference model is supplied—output consistency.
Quick example
anomavision validate `
--model distributions\\padim\\bottle\\anomav_exp\\model.onnx `
--config config.yml `
--runs 20
Typical output includes:
AnomaVision Deployment Validation
────────────────────────────────────
Model: distributions\padim\bottle\anomav_exp\model.onnx
✓ File Exists
✓ Onnx Valid
✓ Onnxruntime Inference
✓ Static Input Shape
Latency: 11.020 ms
FPS: 90.75
....
Supported deployment artifacts include PyTorch, TorchScript, ONNX, TensorRT, OpenVINO, Hailo HEF, and Vitis AI/KV260 XModel.
Note: backend availability means the required runtime components are installed. It does not by itself prove that a specific model has been tested on physical target hardware.
See the Deployment Validation guide for all options, consistency validation, JSON output, backend details, and recommended production usage.
📡 Production data drift
Production data-drift monitoring helps detect changes in the data seen by an anomaly detector after deployment.
It compares production representations against a trusted normal reference population using a rolling window. It can report metrics such as PSI, mean shift, standard-deviation shift, cosine shift, and drift score, with the result available through a dashboard and JSON status file.
The monitor is an observer only: it does not change anomaly scores or localization. Drift is an early-warning signal that should be investigated for causes such as lighting, camera position, preprocessing, or product changes.
Quick example
anomavision detect `
--config config.yml `
--model model.onnx `
--enable-drift-monitoring `
--drift-reference .\\drift\\reference_embeddings.npy
See the Production Data Drift guide for reference generation, metrics, dashboard, configuration, and troubleshooting.
📊 Benchmarks
AnomaVision includes a reproducible benchmark workflow for comparing anomaly-detection performance and runtime characteristics.
The benchmark reports Image AUROC, Pixel AUROC, latency, P95 latency, FPS, model/artifact size, and memory usage. The comparison script uses a shared evaluation contract so results can be reproduced consistently.
⚠️ Anomalib version
The benchmark depends on Anomalib APIs that change frequently. For reproducible results, pin Anomalib to the tested commit:
git clone https://github.com/open-edge-platform/anomalib.git
cd anomalib
git checkout 4207f163a556c5dbf59ed58ed372d8c47dc526a0
Do not run the benchmark against the latest Anomalib main branch unless the integration has been updated and validated for that API version.
Quick example
Run a single class:
python scripts/benchmarks/compare_with_anomalib.py `
--dataset_path D:/01-DATA/Visa_pytorch `
--class_name bottle `
--algorithms padim `
--device cpu
Run all available MVTec classes:
python scripts/benchmarks/compare_with_anomalib.py `
--dataset_path D:/01-DATA/Visa_pytorch `
--all_classes `
--algorithms padim `
--device cpu
With --all_classes, the script automatically discovers supported class directories under the dataset path and runs the same AnomaVision-vs-Anomalib comparison for each available class. --class_name and --all_classes are mutually exclusive. Results are saved under benchmark_results/<backbone>/<algorithm>/<image_size>/<class_name>/.
Historical benchmark results are available for MVTec AD and VisA, including per-class results and visual comparisons. These results are retained for reference; rerun the corrected benchmark before making current performance claims.
See the Benchmark guide for the methodology, commands, metrics, Anomalib version/commit, and detailed results.
⚡ KV260 support
AnomaVision supports a Vitis AI workflow for PaDiM and PatchCore on the AMD/Xilinx Kria KV260. XModel compilation has been validated in the Vitis AI environment; final on-device validation requires the physical hardware.
📚 Documentation
📄 License
AnomaVision is released under the MIT License. See LICENSE.
评论