DESCRIPTOR LEARNING METHOD FOR THE DETECTION AND LOCATION OF OBJECTS IN A VIDEO

摘要

The invention relates to a method for learning class descriptors for the detection and automatic location of objects in a video, each object belonging to a class of objects in a set of classes, the method using: - a learning database, consisting of reference videos and containing annotated frames, each frame having one or more tags identifying each object detected in the frames, - descriptors associated with the tags and previously learned by a preprocessing neural network from the annotated frames of the learning database, - a neural network architecture defined by parameters centralised on a plurality of parameter servers, and - a plurality of computing units working in parallel; in this method, for each class of objects, one of the neural networks in the architecture is trained using as input data the descriptors and tags to define class descriptors; each computing unit uses, in order to calculate the class descriptors, a version of the parameters of the parameter server to which the unit is linked and returns the updated parameters to said parameter server at the end of its calculation; and the parameter servers mutually exchange the parameters of each computing unit to train the neural networks for each class descriptor.