I'm trying to do shot boundary detection using SVM's.

I'm differentiating between 3 shots, cut, fade and dissolve + (normal frames).

I had used two features, shannon's entropy and variance, it gave me 40/635 boundary frames correctly predicted. That is poor accuracy.

What features can I use to improve this?

Training videos: 2 videos from TRECVID (~50000 frames) Test Videos: 1 Video TRECVID (~13000 frames)


So one must be careful in choosing features.

So I fixed a problem that was occuring, if anyone comes across this post, then the above features are okay and will give you pretty good results for CUT detection.

  • $\begingroup$ you might need to work on your feature vectors. I'm pretty sure you can do better than those two single numbers. $\endgroup$ Apr 19 '17 at 13:09
  • $\begingroup$ Check my edit @MarcusMüller, Can you suggest some features or a feature selection method in OpenCV? Adaboost is common but i think get_weak_features is decrepated in newer versions $\endgroup$
    – astroman
    Apr 19 '17 at 13:20
  • $\begingroup$ hm, those are still onedimensional measures for the difference between two frames, right? $\endgroup$ Apr 19 '17 at 13:22
  • $\begingroup$ Yes, also hold on let me get some values for you, check edit again for values $\endgroup$
    – astroman
    Apr 19 '17 at 13:24
  • $\begingroup$ Also a doubt I have been having is should I use longer features (~100 sized feat vector), also check my edit again I added another query $\endgroup$
    – astroman
    Apr 19 '17 at 13:37

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.