Two simple questions. about anomalydetectioncvpr2018

[MubarkLa]

Hi, Thanks a lot for your kind reply, which is very helpful for me.

To my understanding, for one test video, you have 32 segments. The probability of each of the 32 segments to be anomalous is computed via the 16-frame clips inside each segment. May I ask whether my understanding is right? Then I have two further questions.

1. What if the number of the frames in a segment cannot be divided by 16? And what if the number of the frames in a video cannot be divided by 32?

2.For example, given a video that has 640 frames, thus 20 frames for each segment. The 1st segment (frame NO.1〜NO.20) is recognized to be anomalous by your approach. And the ground truth labels that only frame NO.1~NO.19 is anomalous. Then when computing frame-level ROC/AUC, the NO.20 frame will be a false alarm. ➡️ Is it right?

As a green hand, my questions may take you some precious time. I would be very appreciate if you can help with them.

[WaqasSultani]

Hi,

1.After computing abnormality segment wise, I got the probability of each of 32 segments of test video to be anomalous. This means we get a 32-dimensional probability vector for each video. Since we know the frames numbers corresponding to each of the segment, we assign the probabilistic score to each of the frames of the corresponding temporal segment. Finally, ROC/AUC is computed at Frame level. Please let me know if you have some other question.
2. You can use Keras with backend theano