-
Summary of a video
-
Existing methods
- Unsupervised vs Supervised
- Existing methods treat tasks as independent
-
Approach
- First, segment the video into smaller parts
- It is not efficient to isolate every frame, so we oversegment for efficiency
- Next, measure the importance of each frame using an “energy” measure
- Dissimilarity energy vs representativeness energy
- Dissimilarity energy is higher when the frame is different from the previous and next frames
- Representativeness energy is higher when the frame is similar to the previous and next frames
- The sum of these two energies forms a U-shaped curve of similarity
- Finally, select the frames with high energy as the summary
- First, segment the video into smaller parts
-
Evaluation method
- Since it is subjective, quantitative evaluation is difficult
- There is a dataset called the Open Video Project
- There are also datasets like SumMe and Tour20, which consist of summaries and the original videos