|
Initial Results of Autonomous Production and Summarization
|
 
|
| |
|
This page gives some initial results generated by our first prototype of autonomous system for personalized production and summarization of basketball videos. |
| |
|
| Whole quarter production demonstrated at IBC: |
|
IBC demonstration |
|
|
| Explanation: Results of autonomous production of the whole second quarter based on fully automatically detected players. Note that this is a low quality version produced for streaming, which includes blocky artifacts due to high compression rate.
A high quality version is downloadable here. |
|
| 1. Produced Videos under Different Display Resolutions |
|
| Resolution 160x120 |
|
Resolution 640x480 |
 |
| Explanations: In this demo video, we just show both the selected camera/viewpoint sequence and their position in the original view. Other candidates not selected are also displayed for a better comparison, which are shown in the bottom line. Especially, we give results on three display resolutions. From this demo, we could view the movement of viewpoints and their behaviors under different display resolutions. |
|
| |
|
| 2. Results on Camera-wise Viewpoint Selection |
Download Demo Video |
 |
Explanations: In the demo video, the topmost row shows source videos from seven different cameras. We give camera-wise viewpoints individually selected by our criterion in the second, third and fourth row for three display resolutions 160x120, 320x240 and 640x480, respectively. A major observation is that a larger viewpoint will be selected under a higher display resolution. |
|
| |
|
| 3. Results on Smoothing Viewpoint Sequence |
Download Demo Video |
 |
| Explanations: Besides the source views on the topmost row, we compare the smoothed viewpoint sequences under five different smoothing strengths, from a weak setting (sigma2/sigma1=1) to a much stronger one (sigma2/sigma1=16) in five separated views. We could confirm that without smoothing, the flicking visual artifact is very obvious, while the result after proper smoothing significantly improves the perceptual comfort. |
|
| |
|
| 4. Results on Smoothing Camera Sequence |
Download Demo Video |
 |
| Explanations: In this demo, we compared smoothed camera sequences under different camera smoothing strengths, from the weakest one with gamma=0.0 to a very strong one with gamma=1.0. We could confirm that without smoothing, the flicking visual artifact is very obvious due to dramatic camera switching, while the result after proper smoothing significantly improves the perceptual comfort. |
|
| |
|
| 5. Results on Viewpoint/Camera Sequence under Different Display Resolutions |
Download Demo Video |
 |
Explanations: In this demo, we compared the results under three different display resolutions. We display these results in their result display size, and put them together for a better overview.The left graph shows the generated viewpoint sequences, while the right graph shows the generated camera sequences. We confirm that under different display resolutions, our methods could provide different viewpoints which enables us to have a clear view adapted to our limited display resolution. |
|
| |
|
| 6. Results on Generated Sequences under Different User-specified Cameras |
Download Demo Video |
 |
| Explanations: In this demo video, we show that our system could produce different contents to satisfy narrative user-preference, such as user's favorite camera view. For better comparison, we give seven sequences, each of which has one camera view specified, along with the result without any special preference on cameras. For each sub-picture, we give the index of its selected camera and the position of corresponding viewpoint. We see from the demo video that our system produces the closest view to the user-specified one while keeping the whole story fluent. Here, we use a weak setting (sigma2/sigma1=4) for viewpoint smoothing and display resolution is set to 640x480. |
|
| |
|
| 7. Results on Generated Summary under Different Length Constraints |
Download Demo Video |
 |
| Explanations: In this demo video, we put several summaries generated under different length constraints into one screen to compare their contents. If a frame is not selected in a summary, the corresponding area for this summary will be black. In the centeral bottom line, a graph shows the selection status of each frame in those summaries. We could confirm that our summarization system could extract important events. Basically, our system implements video-skimming, but our system prefers segments of long length and averagely high intersts, we could filter those short segments with a very sharp. This is an advantage of our system over conventional video-skimming methods. |
|
| |
|
| Note: All downloadable demo videos are encoded in xvid mpeg 4 format. If you could not replay the demo video, please download the latest vlc media player for your os or other proper players. |
HP of VLC Media Player |
| |
|
