Violent Scenes Detection

Violent Scenes Detection – automatically detects violent scenes in movies

Target users and customers

All content providers may be interested in the violent scenes detection technology to manage efficiently their content database. In particular, the violent scenes detection technology is an interesting feature for VOD services, as it may help users select a movie suitable for the entire family.

Application sectors

Content structuring & browsing: Knowing where the most violent scenes are in a movie gives direct access to these events, thus enabling a non linear browsing in the document.
Content retrieval: By adding additional metadata to contents (i.e., timecodes of violent scenes), the module finds a direct application in the context of content retrieval in videos databases.

Description:

The Violent Scenes Detection module is a system which detects automatically the most violent scenes in a movie. Provided with a set of pre-computed audio and video features from a video stream, the system will output a file with the corresponding violent scenes timecodes. The chosen pre-computed features were proved to be discriminative for the task of violent scenes detection. They are used as an input for a classification sub-system based on Bayesian networks. This sub-system was trained and its parameters learned offline, on a movie database.

The Bayesian model structure was also automatically learned, without using any expert knowledge. The structure learning property enables one to automatically train another model on another type of videos to detect some other type of events, without building the structure of the model manually.

Technical requirements:

The Violent Scenes Detection module takes a set of pre-computed features from a video stream as input.
Currently running under matlab.

Conditions for access and use:

Module currently available as a matlab prototype only.
Corresponding deliverables are all stated QI – i.e., this module is only available to Quaero partners, on their request. Related IPL is proprietary of Technicolor.