Global Mean Square Error Separation Loss
Yun Hou · Journal of Physics Conference Series · 2022
Class imbalance problem greatly affects the accuracy of deep-learning-based object detectors. To weaken its influence, this paper proposes a global mean square error separation loss (GMSES Loss). GMSES Loss assigns a factor to each example, and then enhances the learning of hard-to-learn examples by reducing the factors of easy-to-learn examples. Meanwhile, the global mean square error separation method is introduced to enhance the learning of foreground classes. To verify the effectiveness of our algorithm, experiments are conducted with YOLOv4 and RetinaNet as the baselines, respectively. The results have shown that our loss function can improve the performance of one-stage object detectors without adding any extra hyperparameters.