A problem about the generated anomaly score

[Hepta-Col]

Hi, I am training/testing your model on a fight detection dataset, and I found a strange behavior:


Here are two examples. Red curves are output scores, and green curves are ground truths. So for an anomaly video (i.e. contains a fight event), there are always peaks at the start and the end of the video, no matter whether there is a fight happening at those points or not. I consider this to be closely relevant to your methods to calculate anomaly scores. Could you please help me explain this? Thanks!