You know how it is.
We are always trying to change something up in our video analysis process. That helps us stay in touch with recent technology development and offer the best possible experience to our clients.
Recently, we have tried a couple of things with trailer generation. Let’s see how we make our trailers even better now.
The deal with automated trailers
When it comes to making trailers for any type of video content, it’s no “one-size-fits-all” deal. There are many variables that must be reflected in the trailer for it to fulfill its purpose.
Different genres of movies require different trailers that would facilitate the plot, tease the right scenes, convey the meaning of the content, and attract the specific audience.
You might want to have a trailer made that would explain the plot, focus on action scenes, or feature main cast members.
To back all of that goodness up, you need not only a strong AI algorithm to analyze your original content but also have a good logic behind that algorithm.
AI-generated trailers: how we made them before
Previously, the AIHunters team has used an algorithm that would make the trailers for all kinds of video content with equal distribution in mind. Here’s a breakdown of the process:
- CognitiveMill™ breaks down the footage into separate shots;
- Then it studies those shots, tagging each of them as focused on story flow, packed with action, suspenseful, etc.;
- The software goes on and picks out the shots sequence that is most fitting to the scenario set by the user. If they set the tool to get the action-packed trailer, they get the action-packed scenes;
- Those scenes are then cut and edited together into a cohesive trailer, making sure it picks out the whole variety of environments, cast members, etc.
That works pretty well for producing automated trailers for different purposes.
Think you would have use for automated trailers? Talk to us!
But what if we need a bit more than that? What if we could go for a bit more dynamic scenario where the trailer has different footage throughout the timeline?
For instance, we can have a trailer show a bit of action, demonstrate the cast members, and lay the expositional foundation for the movie or a TV show.
You won’t believe it, but we asked ourselves the same question.
Turns out we could do it — all it took was a bit of tweaking.
New approach to automated trailer generation
The team got their hands on a Netflix series Fauda and tried out the new approach on it. The thing that was new for the engineers was the volume of content for analysis: a season of Fauda comprised 12 episodes, which make up 12 straight hours of footage.
It was interesting to see how fast CognitiveMill™ could chew through that many hours in a single analysis session.
This time, the AI engineers decided to make the software look for scenes that would reflect several aspects of the footage, capturing the essence of a TV show or movie. If the algorithm places shots with those aspects at the right spots on the timeline, it gets rewarded.
Here are those aspects that the algorithm looks for:
- Observation — the story, major plot points, and environments;
- Characters — featuring main characters that appear in the movie or TV show;
- Action — including the most appealing scenes.
What have we got as a result? A well-rounded trailer that in some cases even features the same scenes as the original trailer from Netflix, which was made by humans.
Come take a look: