Method for analysing media content

An apparatus comprising at least one processor, an input/output (I/O) circuit and memory including computer program code, the memory and the computer program code configured to, with the at least one processor, cause the apparatus to at least:

• input a media content object via the I/O circuit into a feature extractor comprising a plurality of semantic layers forming a hierarchy of increasing semantic abstraction levels;
• extract, by the feature extractor, a plurality of feature maps from said media content object on each of said plurality of semantic layers;
• select at least a portion of the media content object to be analysed;
• analyse said plurality of feature maps extracted on each of said plurality of semantic layers from at least the selected portion of the media content object;
• determine, by a feature attention modelling unit, based on the analysis of said plurality of feature maps extracted on each of said plurality of semantic layers from at least the portion of the media content object and an analysis of a previous state of a recognition unit, one or more feature maps selected from said plurality of feature maps of said plurality of semantic layers;
• determine a weight for each of said one or more feature maps;
• repeat the determining steps for different selected portions of the media content object, each time:
• process, based on the analysis of said plurality of the feature maps and of the previous state of the recognition unit, each of said one or more feature maps by applying the corresponding weight,
• input said processed feature maps in an order corresponding to processing said feature maps to the recognition unit, and
• analyse a number of said processed features maps in said order in the recognition unit,
• until an object classification prediction about at least said portion of the media content object is output.

A computer program embodied on a non-transitory computer readable medium, the computer program comprising instructions causing, when executed on at least one processor, at least one apparatus to:

• input a media content object via an input/output (I/O) circuit into a feature extractor comprising a plurality of semantic layers forming a hierarchy of increasing semantic abstraction levels;
• extract, by the feature extractor, a plurality of feature maps from said media content object on each of said plurality of semantic layers;
• select at least a portion of the media content object to be analysed;
• analyse said plurality of feature maps extracted on each of said plurality of semantic layers from at least the selected portion of the media content object;
• determine, by a feature attention modelling unit, based on the analysis of said plurality of feature maps extracted on each of said plurality of semantic layers from at least the portion of the media content object and an analysis of a previous state of a recognition unit, one or more feature maps selected from said plurality of feature maps of said plurality of semantic layers;
• determine a weight for each of said one or more feature maps;
• repeat the determining steps for different selected portions of the media content object, each time:
• process, based on the analysis of said plurality of the feature maps and of the previous state of the recognition unit, each of said one or more feature maps by applying the corresponding weight,
• input said processed feature maps in an order corresponding to processing said feature maps to the recognition unit, and
• analyse a number of said processed features maps in said order in the recognition unit,
• until an object classification prediction about at least said portion of the media content object is output.