A Player Detection Method Based on Multiple Instance Learning

Sports video contains many different human bodies, in which players’ behavior is directly related to the game and sport contents, therefore, accurate player detection is the key issues of sports video analysis. Existing human detection algorithms have achieved good performance on general human detection tasks. However, these methods will detect player and non-player in sports videos at the same time and cannot further distinguish players from other human bodies. It is possible to train a specific model for player detection. However, it needs large amount of bounding box labels for players. This paper proposes an object detection method based on multiple instance learning. After the object detection for all kinds of human body, a multiple instance learning model is designed. It automatically obtains a feature mapping matrix in the way of weakly supervised learning from the training set with image-level labels. Finally, the player and non-player could be distinguished based on the similarity in the mapping feature space. The compared experimental results demonstrate that the proposed could utilized the existed human body detection framework effectively and detect the player with the minimum labeling cost. And the accuracy is same as the model specifically trained with the player bounding boxes.