Real-time Application

[mnik17]

Hello,

I went through the paper and I ended up having a few questions regarding this approach.

1. In the paper you say that you divide the testing videos in 32 parts and then run them through your evaluation (if I understand correctly). Since I need something that would run on a live camera stream, my question would be if you think it's possible to adapt the code in such a way that I'm able to feed the stream directly into the code and then evaluate a certain amount of time (for example 5 seconds) as one segment. In this manner I'd still be able to evaluate everything after a bit more than a minute. And what happens if you use fewer segments, e.g. 10? Is the performance worse and if so, do you have any idea how much worse it is? Because if I have fewer segments, theoretically I'd be able to evaluate and reach a conclusion faster.

2. I'd also be thankful if you can share what kind of hardware are you using for this, as well as what was the runtime, since I'd only be using a CPU and I was planning on optimizing this method in order to run it on it. I was hopeful it will work, because compared to other methods, it seems a bit more lightweight.

Thank you for reading!

[qandeelabbassi]

@mnik17 did you discover the answers to the above questions yet?
@WaqasSultani I have the same questions. Can you please guide us?

[ekosman]

Find the implementation of the paper in Pytorch (contains an online version)
https://github.com/ekosman/AnomalyDetectionCVPR2018-Pytorch