1. Using Cross-Media Correlation for Scene Detection in Travel Videos

2. Outline  Introduction  Approach  Experiments  Conclusion

3. Introduction Why Use Cross Media Correlation for Scene Detection in Travel Video?? What Correlation between photos and video? More and more people get used to record daily life and travel experience both by Digital Cameras and Camcorders. (much lower cost in Camera and Camcorders)

4. Why Use Cross Media Correlation for Scene Detection in Travel Video?? What Correlation between photos and video? People often capture travel experience by still Camera and Camcorders. The content stored in photos and video contain similar information. Such as Landmark, Human’s Face. Massive home videos captured in uncontrolled environments, such as overexposure/underexposure and hand shaking.

5. Why Use Cross Media Correlation for Scene Detection in Travel Video??  It’s Hard for direct scene detection in video.  High correlation between photo and video.  Photo obtain high quality data (scene detection is more easier).

6. Approach  What’s different purpose that people use photo and video even capture same things?  Photo To obtain high quality data, capture famous landmark or human’s face  Video To Capture evolution of an event Utilize the correlation so that we can succeed the works that are harder to be conducted in videos, but easier to be done in photos

7. FrameWork  To perform scene detection in photos: First we cluster photo by checking time information.  To perform scene detection in videos: First we extract several keyframe for each video shot, and find the optimal matching between photo and keyframe sequences

8. The idea of scene detection based on cross media alignment

9. The proposed cross-media scene detection framework Photos Time-based clustering Visual word representation DP-based Matching Videos Shot change detection Keyframe extraction Filtering (motion blur cease ) Visual word representation Scene boundaries This process not only reduces the time of cross-media matching, but also eliminates the influence of bad-quality image

10. Preprocessing  Scene Detection for Photos utilize different shooting time to cluster photo denote the time difference between the ith photo and the (i+1)-th photo as g i g i ＝ t i+1 － t i A scene change is claimed to occur between the nth and (n+1)-th photos. We set K as 17 and set d as 10 in this work. K is an empirical threshhold D is the size of sliding window

11. Preprocessing  Use Global k-means algorithm to extract Keyframe  Detect and Filtering blur Keyframe. It’s no only reduces the time of cross-media matching, but also eliminates the influence of bad-quality images.

12. Visual Word Representation  Apply the difference-of-Gaussian(DoG) detector to detect feature points in keyframes and photos  Use SIFT(Scale-Invariant Feature Transform) to describe each point as a 128-dimensional feature vector.  SIFT-based feature vectors are clustered by a k- means algorithm, and feature points in the same cluster are claimed to belong to the same visual word

13. Visual Word Representation KeyFrames, Photos SIFT Feature point (Feature vector) K-means Visual Word

14. Visual Word Histogram Matching X i denote the i th prefix of X, i.e., X i ＝ LCS(X i,Y j ) denotes the length of the longest common subsequence between X i and Y j

15. Evaluation Data

16. Evaluation Metric The first term indicates the fraction of the current evaluated scene, and the second term indicates how much a given scene is split into smaller scenes. The purity value ranges from 0 to 1. Larger purity value means that the result is closer to the ground truth τ(s i,s j * ) is the length of overlap between the scene s i and s j * τ(s i ) is the length of the scene s i T is the total length of all scenes

17. Performance in terms of purity based on different numbers of visual words, with different similarity thresholds

18. Performance based on four different scene detection approaches Hue Saturation Value

19. Conclusion For video, extract keyframe by global k- means algo. (Scen spot can be easily determined by time information of photos) Representing keyframes and photo set by a sequence of visual word. Transform scene detection into a sequence matching algo.

20. Conclusion  By using a dynamic programming approach, find optimal matching between two sequence, determine video scene boundaries with the help of photo scene boundaries. By experiment on different travel video, different parameter settings, result shows that using correlation between different modalities is effective