Academia.eduAcademia.edu

Figure 5 – uploaded by Kerem Caglar

See full PDFdownloadDownload figure
Figure 4: Key-Frames (top) and Scene Frames by NNE (bottom-up) and MST (bottom-down).

Figure 4 Key-Frames (top) and Scene Frames by NNE (bottom-up) and MST (bottom-down).

Related Figures (4)

Every time a branch is broken and thus a new cluster is created, the nodes inside the cluster can be assumed to show high similarity than the nodes outside the cluster. Knowing that those nodes represent the key- frames, in this way we grouped all the key-frames that are similar to each other. Once a sufficient number of clusters are achieved, it is straightforward to choose one or more key-frame(s) as the scene frame(s), which can represent the scene (cluster). Therefore, for efficient SD, it is extremely important to know where to stop clustering. Usually in supervised methods, a preset threshold value for branch weight comparison is experimentally found and thereof applied for the decision whether or not to break the branch. This might work on some video sequences but obviously not guaranteed to work on every video sequence.
Every time a branch is broken and thus a new cluster is created, the nodes inside the cluster can be assumed to show high similarity than the nodes outside the cluster. Knowing that those nodes represent the key- frames, in this way we grouped all the key-frames that are similar to each other. Once a sufficient number of clusters are achieved, it is straightforward to choose one or more key-frame(s) as the scene frame(s), which can represent the scene (cluster). Therefore, for efficient SD, it is extremely important to know where to stop clustering. Usually in supervised methods, a preset threshold value for branch weight comparison is experimentally found and thereof applied for the decision whether or not to break the branch. This might work on some video sequences but obviously not guaranteed to work on every video sequence.
As clearly seen in this equation, big weight values affect the threshold value more significantly than the smaller ones. Furthermore, the weight values around the mean (in a window of variance) have no effect at all on the threshold value. As a result we can summarize the algorithm as follows:
As clearly seen in this equation, big weight values affect the threshold value more significantly than the smaller ones. Furthermore, the weight values around the mean (in a window of variance) have no effect at all on the threshold value. As a result we can summarize the algorithm as follows:
In ideal case the key-frames, which belong to the same scene should be similar to each other, so gives close distance values but rather different than the key- frames of the other scenes, thus results in large distance values. In this way once the elimination of the close (intra-scene) key-frames are completed, one can expect a region that no elimination is possible until the inter- scene frame eliminations occur. Since the number of scene frames is the difference between total number of key-frames and number of eliminated key-frames, such a region can be directly detected from the number of scene frames vs. threshold curve. In ideal case we can expect a zero-gradient plateau in such a curve but to have a robust algorithm we try to detect the lowest gradient section with the largest width as shown in
In ideal case the key-frames, which belong to the same scene should be similar to each other, so gives close distance values but rather different than the key- frames of the other scenes, thus results in large distance values. In this way once the elimination of the close (intra-scene) key-frames are completed, one can expect a region that no elimination is possible until the inter- scene frame eliminations occur. Since the number of scene frames is the difference between total number of key-frames and number of eliminated key-frames, such a region can be directly detected from the number of scene frames vs. threshold curve. In ideal case we can expect a zero-gradient plateau in such a curve but to have a robust algorithm we try to detect the lowest gradient section with the largest width as shown in
Figure 3: Key-Frames (top) and Scene Frames by NNE (bottom-up) and MST (bottom-down).
Figure 3: Key-Frames (top) and Scene Frames by NNE (bottom-up) and MST (bottom-down).
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved