In this section, we first evaluate the performance of the semantic model using its applications. We have conducted our experiments on twelve archive films and twenty video song clips totally summing up to more than 30 hours of video.
Evaluation of Semantic Hierarchical Model
In this section we evaluate the efficiency of semantic videon model through the applications proposed by us.
| Movies | Videon model | Logical Turning Point | Suggest a Video | Video Classification |
| A Walk to Remember | 13:27 | 00:00:06 | 00:00:38 | 00:00:43 |
| Gone in Sixty Seconds | 13:59 | 00:00:12 | 00:00:49 | 00:00:46 |
| Good Will Hunting | 09:46 | 00:00:12 | 00:00:53 | 00:00:46 |
| Ice Age | 13:04 | 00:00:09 | 00:00:49 | 00:00:38 |
| Ice Age: The Meltdown | 16:02 | 00:00:10 | 00:00:51 | 00:00:44 |
| Infernal Affairs | 11:40 | 00:00:12 | 00:00:52 | 00:00:49 |
| Magnolia | 11:41 | 00:00:14 | 00:01:03 | 00:01:04 |
| Minority Report | 17:56 | 00:00:16 | 00:00:59 | 00:00:59 |
| Mission Impossible | 13:46 | 00:00:12 | 00:00:57 | 00:00:54 |
| Mission Impossible II | 16:36 | 00:00:16 | 00:00:50 | 00:00:43 |
| Mission Impossible III | 13:12 | 00:00:18 | 00:00:57 | 00:00:55 |
| My Sassy Girl | 13:42 | 00:00:10 | 00:00:50 | 00:00:48 |
| The Departed | 08:17 | 00:00:09 | 00:00:47 | 00:00:47 |
In the above Table, we have presented the time taken for model computation (indexing time) and the time taken by applications using this model. The computation time of the model can be compared to the indexing process, which is one-time activity. The average time of building model is around 13 hours per motion picture. However the average computation time of the applications designed on top of this model is less than a minute which is 160 times faster than the state-of-art system.
Logical Turning Point Detection
Generally in a movie, there are three logical parts, where the characters are introduced, then they get together and then some serious stuff happens towards the end. In some moves, the number can deviate a bit. Based on this, the manual annotation of the logical turning points were done.
Our algorithm detects such turning points by looking for object and motion discontinuity. We achieved precision of 83$ and recall of 92% in detecting logical turning points. The results are presented in the above Figure.
Suggest a video
In this section we present video suggestions results for our data set. In the suggestions, movie from same category and sequels were also suggested.
| Movie | Suggested Movie |
| A Walk to Remember | Good Will Hunting |
| Gone in Sixty Seconds | Mission Impossible |
| Ice Age | Ice Age: The Meltdown |
| Mission Impossible | Mission Impossible II |
| Mission Impossible II | Ice Age: The Meltdown |
Remix candidate search
We conducted experiments on 20 song clips, where the aim is to find best remix candidates. Our algorithm found three pairs of songs which are best candidates for remix in the given set. The remixes are generated by mixing the popular song's audio with other songs video.
The remix generated from song videos are presented to the various type of users who have seen and not seen the original video. The user were requested to rank the remix with one of the following ranks
- excellent
- pleasant
- decent
- satisfactory
- unsatisfactory
- unacceptable
The individual ratings of the generated remix are presented below. We have got an average rank of 0.83 out of 5 for the three remixes generated. This rank is inbetween pleasant and excellent.
| Remixes | Computation Time (hh:mm) | Rank (out of 5) |
| I'm a barbie girl | 01:29 | 0.49 |
| Something Something | 01:42 | 1.28 |
| Snegithanae | 01:24 | 0.72 |
Video Classification
For classifying videos into various category, we have given one movie from each category and chosen the movies with have 87% of the maximum score. The result for classifying movies into categories drama and action is shown in the Figure below. In the result, misclassification rate is low, however the algorithm was unable to find category for three movies out of thirteen movies.
