Skip to content

Outlier Detection

Stage 1 of the pipeline: Detecting artifacts and invalid samples in PLR signals.

Available Methods

Ground Truth

Method Description
pupil-gt Human-annotated ground truth masks

Foundation Models

Method Description
MOMENT-gt-finetune MOMENT model finetuned on ground truth
MOMENT-gt-zeroshot MOMENT model zero-shot
MOMENT-orig-finetune MOMENT finetuned on original data
UniTS-gt-finetune UniTS model finetuned on ground truth
UniTS-orig-finetune UniTS finetuned on original data
UniTS-orig-zeroshot UniTS zero-shot
TimesNet-gt TimesNet on ground truth
TimesNet-orig TimesNet on original data

Traditional Methods

Method Description
LOF Local Outlier Factor
OneClassSVM One-Class SVM
SubPCA Subspace PCA
PROPHET Facebook Prophet-based detection

Ensembles

Method Description
ensemble-LOF-MOMENT-... Voting ensemble of multiple methods
ensembleThresholded-... Thresholded ensemble

Configuration

# In Hydra config
outlier_method: MOMENT-gt-finetune

API Reference

flow_anomaly_detection