Gaussian & Mahalanobis distance

just image level detection, no location.

Untitled

  1. fitting a multivariate Gaussian (MVG) per feature level. There are D feature levels.
  2. each gaussian ⇒ one score by Mahalanobis distance.
  3. final anomaly score := non weighted summation of the above scores.

References

Modeling the Distribution of Normal Data in Pre-Trained Deep Features for Anomaly Detection, ICPR20