An improved class-balanced training sample assignment method for object detection

Class imbalance usually exists in the task of object detection based on deep learning, which has attracted extensive attention. When the number of instances belonging to different classes in the dataset is obviously unequal, class imbalance will occur, leading to the object detection model being biased towards over-represented classes during training. To handle the issue of foreground-foreground class imbalance, we design a constraint function for balancing the number of inter-class positive samples, and the improved Class-Balanced Training Sample Assignment (CBTSA) method is therefore proposed in this work. In our method, the quantitative characteristics of various classes in training set are utilized in the constraint function in order to keep the classifier in balance by equalizing the numbers of training positive samples for all kinds of ground-truth boxes. Hungarian algorithm combined with constrained positive sample numbers, CIoU loss and extended cost matrix is then used to calculate the globally optimal positive samples allocation scheme. Experiments on the challenging MS COCO 2017 benchmark are carried out to verify the effectiveness of the method given in this paper. The results demonstrate that CBTSA method boosts the performance of classifier for underrepresented classes and improves the baseline detector on detection accuracy.