Implementation of state-of-the-art video shot boundary detection

Question

I am working on a wide project involving object retrieval from videos.

According to "A Survey on Visual Content-Based Video Indexing and Retrieval", most popular methods are divided in:

• simple "threshold based approach" (global or adaptative)
• supervised learning-based classifiers (with SVM or AdaBoost)
• unsupervised learning-based algorithm (mainly K-Means)

For the moment I implemented on my own a very simple and old-fashioned method based on differences of color-histograms among successive frames.

Nevertheless, I would like to try something more efficient and up to date, without spending too much time, considering that shot boundary detection is not the main topic of my research.

Does anybody know an implementation of an effective algorithm?

Answer 1

I've found the following implementation extremely useful and effective in my own research:

https://github.com/johmathe/Shotdetect

The workhorse of the codebase happens here:

https://github.com/johmathe/Shotdetect/blob/master/src/film.cc#L117-237

It mostly relies on color information for detection of shots.