[FEAT]: Is it possible to add a classification head on efficientad?

[840691168]

Problem Description

When using EfficientAD, I often encounter high false positive rates. In my current use case, many regions with anomaly scores above the threshold are actually not anomalies.

I believe this issue stems from the lack of semantic-level validation after thresholding. Once an area is marked anomalous, there is no further check on whether it truly belongs to the positive class.

Proposed Solution

While EfficientAD is designed as an unsupervised anomaly detector, this proposed classification head is meant to act as an optional post-filtering module in cases where a small amount of labeled data is available. It does not change the core training procedure of EfficientAD, but simply augments its deployment capability in high-precision scenarios.

Alternatives Considered

No response

Additional Context

No response

[samet-akcay]

@alexriedel1, would you have any insights on this?

[alexriedel1]

@840691168 you can sure try this! It's a bit of work involved but you could use the raw outputs of the student and autoencoder as input to a classification head. But don't let the gradients of this extra head flow through the full architecture so that the model doesn't learn to reproduce the anomalous samples..

[samet-akcay]

@840691168, this might be technically possible, and could be integrated. If this is integrated by community, we are happy to merge it to Anomalib.