How to porting the scence detection algorithm to other system?

[willswl]

Is that possible to deal with other videos, rather than YouTube? I'm interested in the scence detection algorithm and want to porting it to deal with my local videos, is that possible? If so, could you teach me more detail steps?
I found there is an opening issue about the installation, but not answers yet, That answer should be also very helpful to me.

[cooperpellaton]

I'll try to post an install script, but until then the following is important:

• This was run on Ubuntu 16.04.
• The server needs to be run with Ubuntu 3.
• The ML needs to run on Python 2.7

You need to have a non-apt-get installed version of OpenCV to work. We now only support the latest version (the latest pull should have migrated this to work with Python3 but I need to clean all the files). This is a good guide to follow for OpenCV install.

To run: just change the parameters in the Highlighter. Most of them are optional, to begin with, but you need to omit the params that deal with a YouTube video and then provide a local file path as the URL.

We have no plans of further development (that I know of) but I'll post a setup script and try to clean the repo when I have some time in the next couple weeks.

[willswl]

Thanks for your kindly answer, I've already working on porting highlighter.py. My working is the same as the steps you replied.

My virtual machine environment is :

• Ubauntu 14.04
• Anaconda python 2.7
  I installed OpenCV with conda - 2.4.10.
  And you are using opencv version of 2.4.x right?
  at the first I installed version 3.1, but it told me 'the object has no attribute BackgroundSubtractorMOG', so I downgraded the version to 2.4.10, and BackgroundSubtractorMOG error resolved.

[cooperpellaton]

Latest version in the repo should be for OpenCV for Python 3.5x. If you're getting errors, the port should be rather trivial because only a couple of methods changed.

[willswl]

Thank you very much. the program finally works on my windows work station, but I'm a little confused about the result: every scene got the same score - 2.0, even I assigned differenct category number to the same video.
I was using "nn_model.pickle" from the respository, is there anything I missed?

More over, is there any guidance on how to set the category for a local videos? What is the category number range?

[willswl]

@alioben @vontell Could you help take a look my question? Eery scene got a 2.0 score . The get_scenes function works well, and I can distill the whole video into several scenes, but the prediction score is always the same. I didn't change the predict code, just reused the existing "nn_model.pickle" modle, as below:
scenes[i]["score"] = np.asscalar(model.predict(vect.reshape(1, -1)))