190 likes | 319 Views
Mining TV Broadcasts 24/7 for Recurring Video Sequences. Jeong , Dongseok. Before the start – Question. There are two techniques used for Video Fingerprinting : CPF(Color Patches Features) and Gradient Histograms. What is the main idea of these techniques?
E N D
Mining TV Broadcasts 24/7 for Recurring Video Sequences Jeong, Dongseok
Before the start – Question • There are two techniques used for Video Fingerprinting : CPF(Color Patches Features) and Gradient Histograms. What is the main idea of these techniques? • What methods are used for similar image searching?
The purpose of the paper • Detecting recurring video clips in a TV broadcast stream • In this case, to detect recurring commercials in a TV broadcast stream • How can we overcome the storage and computing issue of video data?
The structure of the paper • Introduction • Fingerprinting Video Streams ☜ • CPF, Gradient Histograms • Explain own search algorithm ☜ • Experimental results • Related Work • Conclusion
Video Fingerprinting • Color Patches Features(CPF) Each frame is reduced to averages of pixels
Video Fingerprinting • CPF example
Video Fingerprinting • Gradient Histograms Edge-based features
Video Fingerprinting • Gradient Histograms cont.
Video Fingerprinting • Finally, we can measure the distance between two image and • The distance between two sequences , of length L is given by
Searching Algorithm • Finding similar images to the source video • Use inverted index and Locality Sensitive Hashing • Compare short clips from each source • Finding the start-point and end-point of repeated sequences • Classifying the repeated video
Optimizing some parameters • Clip Length : choose 25 frames
Optimize some parameters • Minimum Fraction of Matched Frames • choose 20%
Optimize some parameters • Maximum Number of Entries in Hash Table • choose 100 entries per hash value
Optimize some parameters • Minimum Length of Duplicates • Choose 100 frames
CPF VS GH • Searching for flips with GHs is up to 30% faster than using CPFs • But CPFs are faster to evaluate and need a smaller amount of storage • (a) : Chart TV • (b) : Sky Sports News
Mining Different TV Channels • Apply the system to a variety of broadcast stations
Mining Different TV Channels • The remaining false detections are mainly caused by repeated news stories(in ARD : the German public broadcaster) • Gemini is an Indian TV channel – for non-natives what are commercials and what not?
Question Again • There are two techniques used for Video Fingerprinting : CPF(Color Patches Features) and Gradient Histograms. What is the main idea of these techniques? • What methods are used for similar image searching?
The End! • Any Question?