Hello,
the currently expected evaluation format is unclear to me: Do you require us to normalize the predicted scores to lie in the interval [0, 1], without using/accessing anomalies?
Or are the statements "We expect a single probability score per sample, where 1.0 means anomaly" and "Each pixel's value is a probability score where 1.0 means anomaly." meant to tell us that anomalies are supposed to achieve higher scores than normal data (though some assumptions will have to be made then when converting/normalizing to uint8 value range)?
Regarding F1-Max computation: Is the best threshold allowed to vary between segmentation and anomaly detection task?
Best,
ORippler