Balanced Loss for Accurate Object Detection

Abstract

In object detection, the imbalance problem often occurs when the number of training samples of different categories varies greatly, or multiple loss functions need to be minimized which is harmful to the performance of the detector. In this paper, we consider that the imbalance problem can be implied by the imbalance of gradient distribution. To address these imbalance issues, we analyze the gradient of cross-entropy loss and propose balanced cross-entropy (BLCE) loss and balanced binary cross-entropy (BBCE) loss for solving objective imbalance and class imbalance issues respectively. The BLCE loss significantly reduces the overall classification loss and keeps the classification loss and regression loss balanced. Furthermore, the BBCE loss automatically down-weight the contribution of inliers during training and rapidly focus the model on outliers. Ablation studies on object detection and image classification demonstrate the effectiveness of our loss function. We replace the corresponding losses in Libra R-CNN and evaluate our detector on the COCO test-dev. Our results show that Libra R-CNN can surpass the accuracy of many existing state-of-the-art detectors when training with our balanced loss.