vaas 0.1.6

VAAS (Vision-Attention Anomaly Scoring) is a dual-module vision framework for image anomaly detection and localisation.

VAAS: Vision-Attention Anomaly Scoring

VAAS (Vision-Attention Anomaly Scoring) is a dual-module vision framework for image anomaly detection and localisation.
It combines global attention-based reasoning with patch-level self-consistency analysis to produce interpretable anomaly scores and dense anomaly maps.

This repository provides the inference-ready implementation of VAAS for research engineers and practitioners.

---

Paper

The conference paper link will be added upon publication.

---

Architecture

VAAS integrates two complementary components:

• Fx — Global Attention Module
  A Vision Transformer capturing semantic/global irregularities from attention patterns.

• Px — Patch Consistency Module
  A SegFormer-based model capturing local inconsistencies across image patches.

These combine to produce:

• S_F — global attention fidelity
• S_P — patch-level plausibility
• S_H — hybrid anomaly score (final)

S_H is continuous and reflects anomaly intensity, not just presence.

---

Installation

pip install vaas

Important: VAAS requires PyTorch + torchvision.
Install the correct PyTorch build for your system (CPU / CUDA / ROCm):

https://pytorch.org/get-started/locally/

---

Quick Inference Example

from vaas.inference.pipeline import VAASPipeline
from PIL import Image

pipeline = VAASPipeline.from_pretrained(
    "OBA-Research/vaas-v1-df2023",
    device="cpu",
    alpha=0.5,
)

img = Image.open("image.jpg").convert("RGB")
out = pipeline(img)

print(out["S_H"])

Output format

{
    "S_F": float,
    "S_P": float,
    "S_H": float,
    "anomaly_map": numpy.ndarray  # shape (224, 224)
}

---

Model Variants

Version	Training Data	Description
v1	DF2023 (10%)	Initial public inference release

Future updates will include:

• full DF2023 training
• cross-dataset benchmarking
• improved calibration
• more efficient backbones
• richer visualisation modes

---

Example Notebooks

A set of runnable example notebooks covering:

These cover:

• Quick-start inference with VAAS
• Visualising anomaly maps
• Batch / folder inference
• How S_F, S_P, and S_H are combined
• Device selection (CPU, CUDA, MPS)

If you would like to contribute a notebook, see
CONTRIBUTING.md for guidelines.

---

Roadmap

• Cross-dataset evaluation
• Larger-scale training
• Model compression
• Batch inference and folder-level CLI
• Extended anomaly-map visualisation
• Streamlit/Gradio demo (optional)
• ONNX / TorchScript export

---

Contributing

We welcome contributions that improve the usability, robustness, and extensibility of VAAS.

Please see the full guidelines in CONTRIBUTING.md.

---

Citation

Citation will be added once the paper is published.

---

License

MIT License.

---

Maintainers

OBA-Research
https://github.com/OBA-Research
https://huggingface.co/OBA-Research