index.md

SARIAD

Benchmarking suite for synthetic aperture radar imagery anomaly detection (SARIAD) algorithms. SARIAD integrates SAR datasets, anomaly detection models and pre-processing methods with Anomalib and evaluates them with a common set of metrics (paper).

The flow of SARIAD, adapted from Anomalib. The component lists in the figure date from the paper; see Datasets and Models for what is available now.

Install

git clone --recurse-submodules https://github.com/Advanced-Vision-and-Learning-Lab/SARIAD
cd SARIAD
pip install -e ".[dev]"          # Python >= 3.13

Some models need code from git submodules (SARATR-X, SAR-CNN); --recurse-submodules (or git submodule update --init) fetches them. Importing SARIAD does not require them.

Use

Describe experiments in a YAML file and run them, repeated and compared in a LaTeX table (see SARIAD/config/default.yaml):

sariad --config SARIAD/config/default.yaml --dry-run    # check names and parameters, no downloads
sariad --config SARIAD/config/default.yaml

or use the pieces directly:

from anomalib.engine import Engine
from SARIAD.datasets import SSDD
from SARIAD.models import PadimACE, YOLOAnomaly       # or any anomalib model, e.g. anomalib.models.Padim
from SARIAD.pre_processing import MedianFilter
from SARIAD.utils.inf import Inferencer

datamodule = SSDD()
model = YOLOAnomaly(weights="yolo11n.pt")
engine = Engine()
engine.fit(model=model, datamodule=datamodule)
print(Inferencer().evaluate(model, datamodule, engine=engine))

:maxdepth: 2
:caption: Contents

datasets
models
source/modules