I have trained an unsupervised anomaly detector for surveillance videos. After inference, I rescale the scores between max/min from the resulting scores array.

scores = (scores - min(scores))/max(scores)

Every paper I read has an AUC score describing their model's accuracy. My question is, how do they rescale the scores? The test set consists of 13 videos with ground truth labels. These are some ways I think most of the papers do it.

  1. inference all 13 videos and, after that, rescale the scores between min/max, then calculate AUC
  2. inference one video and rescale the results immediately. In the end, append all the scores and calculate AUC

I want to have a comparable metric; this only works if I do it the "right" scientific way. But I can't find any description of the evaluation process in the papers; they say, "AUC score is X%."

Hopefully, someone understands my question and can help me with my problem.

Thanks :)

  • $\begingroup$ What is the "scores" you use? $\endgroup$
    – lpounng
    Dec 14, 2022 at 9:01
  • $\begingroup$ It is a reconstruction error from the prediction of the model and the ground truth. In my model, it would be frame t+1 (prediction) and frame t+1 we have given. error is then calculated via l2norm(pred - gt) $\endgroup$
    – TecK97
    Dec 14, 2022 at 9:03
  • $\begingroup$ Use a separate (labeled) validation to find all hyperparameters, including scoring parameters $\endgroup$
    – Jon Nordby
    Dec 18, 2022 at 21:09


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.