Clusterization of anime faces, active learning #18
Description
The rational of this task is to improve the current routine that extracts the faces of the characters from the movie. We currently have deciders that split the movie to the series of images, finds face with YOLO. Then, we need to select the faces of the characters we are interested in. Current solution employs WD14Tagger to tag the faces for trairs + logistic regression to predict the character from traits. It can surely be improved.
The improvement would be to employ the solution like this https://github.com/deepghs/imgutils (there are other and often new ones are released, so need to check for state-of-the-art solution before starting). This needs to be implemented as a brainbox decider.
Then, if the quality is suboptimal, an active learning can be employed: the user is presented with some clusters and quickly annotates which faces do not belong to the clusters assigned. Then, the network needs to be fine-tuned with this updated data and the better clusters should be produced. This also needs to be implemented as an endpoint in the brainbox decider.