Weak Supervision in Enemy Detection Based on Computer Game Output Video Stream

Abstract

This work contains a solution for image classification and enemy detection in the output video stream of a computer game. Weak supervision was used to achieve the goal. It shows that an image dataset with a certain number of incorrect classification labels can be used to correctly build a classification model that distinguishes between images containing and not containing an enemy. Based on the results of such classification stage and the use of class activation maps, a method for detecting enemies on positively classified images was proposed. The tedious process of image labeling, which is necessary for supervised learning, does not occur here.