I want to do shot boundary detection via SVM's.

I'm dividing the frame into nxn blocks. Per block I'm finding these features:

shannons entropy edges (H,V,Diag) standard deviation

for consecutive frames and corresponding blocks, the above features are compared via differences and L2 norms, I threshold over that, if above threshold then given a 1 else 0. That is, the feature value belonging to that feature for that particular block becomes 1.

This assignment happens per feature for every block. For a frame, the feature vector is the summation of all the feature vectors of its blocks.

Then I take +-N/2 frames around it and concatenate their feature vectors to create a single feature vector (for that frame).

I trained the SVM using this. There are so many normal frames and only a few shot boundary frames. I'm dividing amongst cut,dissolve, fade, other and normal frames.

I'm unable to properly detect the shot boundaries.

Can someone suggest which features I should use and how I should select them for shot boundary detection? I think I may also have not done quite well on the threshold selection, so any ideas how I could do that better?

I'm using TRECVID's 2001/2002 database.


1 Answer 1


Basically, Shot Boundary detection technique's output is affected by camera motion, object motion and illumination conditions. The feature you choose should be invariant to these issues. The problem in your case may be due to poor training dataset/ imbalanced dataset. threshold can be fixed experimentally and it is tedious.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.