Visual Inspection System for Testing Anomalies: find, localise and report defects in images of manufactured parts and materials.
Status: early development. The pipeline below is the target design; this README tracks what actually works.
Give VISTA an image of a part. It returns an anomaly score for the whole image, the regions it thinks are defective, and a structured inspection record you can store, audit and evaluate against ground truth.
flowchart LR
A[Image] --> B[Preprocess]
B --> C[Anomaly model]
C --> D[Score map]
D --> E[Threshold + region extraction]
E --> F[InspectionResult]
F --> G[(Results store)]
F --> H[Annotated image]
G --> I[Evaluation: AUROC, PRO, latency]
Every stage exchanges data as an InspectionResult (src/vista/schema.py), validated at the boundaries.
- Repo tooling: lint, tests, CI
- Inspection result schema
- Dataset preparation (MVTec AD)
- Baseline anomaly model (PatchCore, see ADR 0001)
- Results store (JSON Lines) and image-level metrics (AUROC, precision/recall at a target false-positive rate)
- Pixel-level evaluation (pixel AUROC, PRO) and latency
- Inference API
- Web demo
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
ruff check . && pytestWe benchmark on MVTec AD, which is licensed CC BY-NC-SA 4.0 (non-commercial). The dataset is not redistributed in this repository; only manifests and preparation scripts are.
Built by V2GWorks.
Code: Apache-2.0. Datasets keep their own licences.