This paper addresses foreground-foreground imbalance in object detection. Firstly, we introduce Mini-batch Stochastic Gradient Descent (MBGD) with YOLO and the foreground-foreground imbalance problem. Then T-distribution is devised and proved to smoothen the imbalanced distribution and allocate at least a representative for each class. Furthermore, Mini-Batch Imbalance Smoothing method (MB-IS) is proposed to address the foreground-foreground imbalance by following T-distribution and proportionally assigning class weights in a mini-batch. Finally, Extensive experiments on our own transaction dataset and VOC2007 dataset demonstrate the superiority of MB-IS with certain mini-batch size.
- Foreground-foreground imbalance
- Mini-batch stochastic gradient descent
- Object detection