Wednesday, January 13, 2010

Shot Detect Performance

We are still having problems with compiling the Shot Detect software. For now, we will use the shot detect executable to get an idea of how well it performs. We ran shot detect on 4 videos from our dataset. Eventually we will implement this performance check on our entire Dataset. The results are as follows...

We define Error as (Undetected Shots + False Positives) / Total # of Shots

Song 1 -
52 shots
Undetected Shots - 5
False Positives - 0
Error - .096

Song 2 -
57 shots
Undetected Shots - 5
False Positives - 0
Error - .087

Song 3 -
49 shots
Undetected Shots - 2
False Positives - 4
Error - .122

Song 4 -
27 shots
Undetected Shots - 10
False Positives - 2
Error - .444

The nature of the first three songs is such that the shot detection is very obvious to the human eye. Song 4 isn't so obvious, i.e. at times, there are effects like dissolve and fade between shots. These sorts of transition between shots are not detected by the shot detect program.


video

No comments:

Post a Comment